Character Consistency in AI Videos Solved - How I make hour-long videos with MiniMax H3
Credibility score: 65/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "Character Consistency in AI Videos Solved - How I make hour-long videos with MiniMax H3" and rated it 65/100 for credibility (a BS score of 35/100 — mostly credible), on 2026-11-20. Its weakest claim — "Claims a total solution to AI consistency via three methods" — scored 45/100 and was flagged as confidence mismatch. 54 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 54 claims analyzed: 0 scored under 40, 19 between 40 and 69, and 35 at 70 or above.
Claims analyzed
Claims a total solution to AI consistency via three methods — Confidence Mismatch (45/100)
At 0:00
Claims to have 'solved' a massive industry struggle with zero proof in the intro 🚩
Why this score: He's positioning himself as a savior of the medium, claiming to have fixed long-standing struggles before even showing a single frame of proof. It's high-confidence posturing without the evidence to back up the 'no more' hype.
Original quote: “Making AI video has long struggled with maintaining consistency across multiple shots or during long clips. Well, no more. I'm here to explain how you can create long-form videos and even multiple shots and maintain faces and outfits in perfect sync. This video covers three levels of methods to…”
Sources: Making Generative AI Output More Consistent | by Ken Arya | Medium, How to Achieve AI Consistency for IP Solutions in 3 Steps | Black Hills AI, The Trinity of Reality: Why Consistency is the Soul of the Next AI Frontier - Neuronad - AI News and AI Tools for Everyone
Using a single prompt to prove randomness in AI — No Frame (75/100)
At 2:06
Demonstrating the lack of control with a real-time example. He's showing, not just telling 👁️
Why this score: He uses a specific example of a prompt and its resulting image to demonstrate how much detail is generated 'randomly' without being explicitly requested, which serves as valid evidence for his argument.
Original quote: “And if I run this right now, I will get and this is with these seeds. I will get this result here. But what you'll notice is that there's nothing in this prompt that says that she has a gray sweater. There's nothing in this prompt saying that there's some jars at the back of the picture...”
Setting up the scene with specific details for AI generation. — No Frame (75/100)
At 2:30
Just describing the setup for the AI, no tricks here. Standard procedure. 😈
Why this score: The speaker is simply outlining the visual parameters they are feeding into the AI for image generation. This is a straightforward description of their process, not a claim that requires external verification or contains manipulative framing.
Original quote: “That there are, you know, three other people here, that she has brown hair, that the coffee is on the left side of her laptop.”
Downplaying seed importance while explaining its core function. — Volume Game (45/100)
At 3:35
Says the seed 'doesn't really matter,' then immediately explains how it directly causes 'different results.' That's not 'doesn't matter,' that's crucial. 💀
Why this score: The speaker initially dismisses the importance of the seed number, stating it 'doesn't really matter.' However, in the very next breath, they explain that the seed 'randomize input' and that 'different seeds' lead to 'different results.' This creates a contradiction where the initial dismissal is undermined by the subsequent explanation of its direct impact on output variation. It's a classic volume game: make a bold, dismissive claim, then quietly walk it back with the truth.
Original quote: “The seed number doesn't really matter because there's other things that affect the generation much more. But what this seed does is randomize input. So if we keep just generating multiple videos with different seeds, we're going to get different results because of that randomness.”
Setting up a demonstration of AI consistency challenges — no trick here. — No Frame (75/100)
At 4:30
Just laying out the premise for the demo. Straightforward, for once. 😈
Why this score: The speaker is clearly stating the purpose of the upcoming demonstration: to show the difficulty in maintaining consistency with AI-generated video characters. There's no hidden agenda or manipulative language here; it's a direct setup for the problem they intend to solve.
Original quote: “as well. But the quality here is not the point. I just want to show you how hard it is to control the consistency if we do this. So, we see now there's a woman.”
Changing a character's ethnicity while claiming 'everything else' stays the same — a setup for a controlled experiment. — No Frame (75/100)
At 4:45
He's setting up a controlled test, changing one variable. Let's see if it holds. 😈
Why this score: The speaker is outlining a specific change to the prompt (ethnicity) while asserting other parameters (seed, resolution) will remain constant. This is a standard approach for demonstrating the impact of a single variable, and there's no immediate indication of manipulative framing. It's a clear statement of intent for the experiment.
Original quote: “So, with this, I'm going to keep everything else in the same. But if I now said, well, this woman kind of looks Asian. Maybe I want a Latina woman. I'm not going to change anything else. seed remains the same resolution.”
Claiming 'everything else' remains the same for a controlled test — setting expectations for the demo. — No Frame (75/100)
At 4:55
He's emphasizing the controlled variables. A good setup, if he actually sticks to it. 🔥
Why this score: The speaker reiterates that only the character's ethnicity is being changed, with the seed and resolution remaining constant. This is crucial for a fair demonstration of consistency. It's a clear statement of the experimental conditions, not a trick.
Original quote: “I'm not going to change anything else. seed remains the same resolution. Everything else is going to remain the same. And now I'm going to hit run again and we'll see how that turns out.”
Acknowledging background changes despite claiming 'everything else' was the same — a slight inconsistency. — Volume Game (45/100)
At 5:07
Said 'everything else' would be the same, then immediately points out background changes. Classic bait and switch. 🚩
Why this score: The speaker initially claimed 'everything else' would remain the same, but then immediately points out 'some changes in the background' and a 'different counter.' This is a subtle contradiction, where a broad claim is made, then quietly walked back with specific exceptions. It's a 'volume game' because the initial strong assertion is undermined by the subsequent details.
Original quote: “And now what you'll see is that this is surprisingly similar to the previous woman. She does look a bit more Latina now. And there are some changes. Let me turn off the sound here. But there's some changes in the background because there's now a different counter here and some other changes.”
Just setting the stage for the solution. No tricks here. — No Frame (75/100)
At 6:30
He's just telling you what's coming. A rare moment of honesty. Enjoy it.
Why this score: The speaker is simply introducing the topic of the video, which is how to fix character consistency issues. There's no hidden agenda or manipulative language here.
Original quote: “I'll show you how to fix it instead.”
Showing the specific workflow in MiniMax H3. Straightforward demonstration. — No Frame (75/100)
At 7:30
Actually showing the steps. A demonstration, not a declaration. Shocking, I know.
Why this score: The speaker is walking through the actual steps of using MiniMax H3's 'reference to video workflow'. This is a direct demonstration of the tool's functionality, not a claim that requires external verification.
Original quote: “So what I'm going to do, I'm going to go into my templates here. I'm going to look for H3 and we're going to use the reference to video workflow.”
Claims MiniMax H3 takes 30 minutes to generate an image, so he uses ChatGPT. That's a bold claim about a tool he's promoting. 🚩 — Confidence Mismatch (45/100)
At 7:46
He's promoting H3, but then says it takes 'half an hour' to generate an image, so he uses ChatGPT. That's not a feature, mortal, that's a flaw. 💀
Why this score: The speaker is promoting MiniMax H3 for video consistency, yet he claims it takes 'half an hour' to generate a single image with it, forcing him to use ChatGPT. This contradicts the implied efficiency of the tool he's showcasing, creating a confidence mismatch between his endorsement and his practical workflow. A top comment also disputes this, suggesting a setting issue.
Original quote: “I'm going to take an image that I made before, which I have here in my folder. And this is an image that I got from chat GPT instead of generating it myself because again it takes half an hour with this H3 model. Plus chat GPT makes good images.”
Speaker adjusts settings for speed over quality, setting up a demonstration. — No Frame (75/100)
At 8:30
Just setting the stage for a demo. No tricks here, just showing the process. 😈
Why this score: The speaker is transparently explaining their workflow choices for the demonstration, prioritizing speed for the sake of showing the process quickly. This is a straightforward explanation of their method.
Original quote: “this. And then we're going to go and add
that as the prompt down here. Generate 5
seconds. Still, I'm going to turn on the
Laura because again, the quality of this”
Speaker explains why current video quality isn't important, promising future detail for consistency. — No Frame (75/100)
At 8:37
Explaining the method, not making a claim. Just showing the gears turning. ⚙️
Why this score: The speaker is outlining their strategy for the demonstration, stating that current quality settings are for speed, and that higher quality will be used later when consistency is the focus. This is a clear, procedural explanation.
Original quote: “video is not important at this point.
And I'll show you when it's going to be
important in a moment because obviously
we want to have details when we want to
look at consistency. And I also want to
drop this down to 0.1 again just because
I want to get something quickly.”
Speaker acknowledges current process is slower than text-to-prompt, setting expectations. — No Frame (75/100)
At 8:54
Setting expectations for the demo's speed. No grand promises, just a heads-up. ⏱️
Why this score: The speaker is managing expectations by admitting that the current method is slower than a direct text-to-prompt approach. This is a straightforward observation about the tool's performance in this specific context.
Original quote: “This is still going to be slower than in
text to prompt, but let's hit run and
we'll see how quickly we get the new
results from the generation.”
Speaker notes the AI generated a 'police woman' despite not explicitly changing the prompt. — Confidence Mismatch (45/100)
At 9:02
Claiming 'didn't change that' when the prompt clearly included 'police woman' earlier. Selective memory, or just a slip. 😈
Why this score: The speaker states 'I didn't change that' regarding the 'police woman' aspect, but earlier in the video (and implied by the subsequent correction), the prompt *did* include 'police woman'. This suggests a slight misremembering or downplaying of the prompt's influence, creating a minor mismatch between the stated action and the likely reality of the prompt's evolution.
Original quote: “And there we go. And now you see that
this image is much closer in appearance
to what we have here. Even if we just
said that this is a Latina woman. Even
said it was a police woman. I didn't
change that.”
Speaker suggests watching the generated video to check for police references. — No Frame (75/100)
At 9:15
Just checking the output. A simple, direct observation. 🧐
Why this score: The speaker is proposing a direct observation of the generated video to verify if the AI included police references. This is a straightforward, verifiable action.
Original quote: “So let's just actually we can watch the
movie first and see if there's actually
some police references in here.”
Speaker explains the importance of setting the seed value for consistency. — No Frame (75/100)
At 9:35
Explaining a technical detail crucial for AI consistency. No spin, just facts. 💡
Why this score: The speaker is explaining a fundamental concept in AI generation (seed value) and its importance for achieving consistency. This is a clear, technical explanation without any manipulative framing.
Original quote: “I'm going
to take out the police for now and get
back to that later. And I also forgot to
set the seed value. And now this is
important because what you'll see now is”
Speaker points out an invented uniform emblem in the AI-generated image. — No Frame (75/100)
At 10:01
Directly pointing out an AI hallucination. Just an observation. 👁️
Why this score: The speaker is making a direct observation about an element (the emblem) that the AI generated but was not present in the reference image. This is a straightforward analysis of the AI's output.
Original quote: “The other thing that's worth noticing
here, look at this symbol here.
This is obviously supposed to be some
kind of uniform emblem or something like”
Running a demo with a 'fixed seed' and removing a 'policewoman reference' to show AI control. — No Frame (75/100)
At 10:30
Just setting up the next demo. He's showing the process, not making a claim yet. 😈
Why this score: The speaker is demonstrating a technical process, explaining the parameters for the next AI generation. This is a straightforward setup for a visual example, not a rhetorical trick.
Original quote: “So, let's run this now with a fixed seed and without the policewoman reference, and we'll see how that turns out. So,”
Stating a 'strong visual reference' allows for 'very, very similar' results even with prompt and seed changes. — Confidence Mismatch (45/100)
At 11:05
He's making a big claim about 'very, very similar' results despite major changes, but the demo hasn't fully proven that yet. 🚩
Why this score: The speaker asserts that a strong visual reference will maintain similarity even with significant changes to the seed and prompt. While this is the goal of the feature, the demonstration is still ongoing, and the 'very, very similar' outcome is presented as a certainty before the full evidence is shown, creating a slight mismatch between confidence and immediate proof.
Original quote: “And even if we change the seed and we change the prompt, because that visual reference is so strong, that allows us to create something that's very, very similar. Even if we now change the prompt to do something else.”
Changing the 'entire prompt' to demonstrate character consistency, calling the process 'movie magic'. — No Frame (75/100)
At 11:18
He's just setting up the next step in the demo, using a bit of flair. It's a direct action. 😈
Why this score: The speaker is clearly explaining the next step in the demonstration, which involves a significant change to the prompt. The phrase 'movie magic' is a colloquialism for the AI's processing, not a deceptive claim.
Original quote: “For example, we can go down here and we'll say that instead of walking to the bar, she will be getting up from her laptop and put on a coat. So now we're going to change the entire prompt. And again, we're going to hit run now and do some movie magic.”
Acknowledging AI's unpredictability: 'you don't always get exactly what you're looking for'. — No Frame (75/100)
At 11:37
He's being honest about the AI's quirks. A rare moment of transparency. 😈
Why this score: The speaker is openly admitting a limitation of AI generation, specifically that it doesn't always produce the exact desired outcome. This is a straightforward and honest observation, not a manipulative framing technique.
Original quote: “And once this completes, now we should see that if we turn off the volume first and we hit play, she will close her laptop. She can close her laptop, but she picks up her coat and then she is possibly leaving afterwards. She's also leaving her laptop and her coffee. I don't know why, but this is…”
Explaining the consistency method: using reference images and descriptions, claiming H3 is 'very good' with multiple references. — No Frame (75/100)
At 12:02
He's laying out the core technique and H3's capability. It's a direct explanation of the workflow. 😈
Why this score: The speaker is summarizing the method for achieving consistency using reference images and describing the scene, and then highlighting H3's specific strength in handling multiple reference images. This is a clear explanation of the process and the tool's features, without any apparent deceptive framing.
Original quote: “So that is essentially the first thing that we can do for our consistency is that we start out with a reference image and then we describe what happens in that reference image and that allows us to create images from any situation and camera angle and then we can start a video from that particular…”
Explaining how H3 uses reference images for character consistency. — No Frame (75/100)
At 12:30
Just laying out the process for using reference images. No tricks here, just a demonstration.
Why this score: The speaker is explaining a feature of the MiniMax H3 model, specifically how it uses multiple reference images to maintain character consistency across different scenes and outfits. This is a straightforward explanation of a technical capability.
Original quote: “additional images now, for example, how her uniform looks or something like that, if she wanted her to be a police woman again, then we could have references that show that.”
Clarifying that reference images aren't limited to animation, but can guide new scenes. — No Frame (75/100)
At 12:48
Still explaining the core functionality of using references. It's a clear, direct statement of capability.
Why this score: The speaker emphasizes that the reference images serve as a guide for character consistency across entirely new scenes, not just as source material for direct animation. This is a key aspect of the H3 model's flexibility being demonstrated.
Original quote: “Then we can have multiple images to basically have those images be animated if you wish. And again, it doesn't need to animate this image either. This can be a reference for a different scene.”
Demonstrating a prompt to place the referenced character in a new, unrelated scene. — No Frame (75/100)
At 13:15
He's showing the prompt in action, how it takes the character from one image and puts her in a completely different scenario. It's a direct example.
Why this score: This is a live demonstration of the H3 model's capability to transfer a character's likeness to a new, distinct scene based on a text prompt, using the previous image as a reference. It directly illustrates the functionality being discussed.
Original quote: “We can say a Latina woman seen in picture one is playing piano at a dark bar with lots of cigar something like this. Now, we're creating a scene that has absolutely nothing to do with this coffee picture here, but we're saying that it's going to be this woman that we want to be in this situation.”
Emphasizing fixed seed for reproducibility, but prioritizing description and visual references. — No Frame (75/100)
At 13:39
He's just explaining a technical detail about AI generation — fixing the seed for consistent results. It's standard practice.
Why this score: The speaker is providing a technical detail about AI image generation: keeping the 'seed' fixed allows for reproducibility. He also correctly notes that for character consistency, the prompt description and visual references are more critical than the seed itself. This is a straightforward technical explanation.
Original quote: “And now we'll hit run again and we'll see what the results are in a few moments. Keep in mind again, we're keeping the seed fixed so that we can recreate our results if we want to do that. But changing the seed isn't necessarily going to be as important as the actual description and the visual…”
Claiming the AI maintained character's appearance, including new clothing not in original reference. — No Frame (75/100)
At 14:00
He's showing the result and pointing out the consistency, even with new elements. It's a direct observation of the output.
Why this score: The speaker is presenting the result of the AI generation, highlighting that the character's likeness is maintained even when new clothing (blue jeans) is introduced that wasn't present in the original reference image. This is a direct observation of the AI's performance, which is the core of the video's topic.
Original quote: “And now we got the video done. 23 seconds. We still see roughly the same woman. Again, we can go into details and then make this much bigger to try this out. But now we see the same shirt. She even has those blue jeans even though they are not in the original image.”
Claiming easy situation changes with single images — then immediately undercutting it. — Volume Game (45/100)
At 14:43
Boasts about flexibility, then immediately says it 'quickly falls apart' for anything over 5 seconds. Classic bait and switch. 😈
Why this score: The speaker presents the ability to put a character into a new situation using a single reference image as a positive, but then immediately qualifies it by saying this method 'quickly falls apart' for longer shots. It's a 'volume game' where a bold claim is made, then quietly retracted or severely limited in the same breath, making the initial claim less impressive upon closer inspection.
Original quote: “But yeah, I can take this image now and then put her into a new situation, even if the original image does not show that situation.”
Admitting single reference images fail for longer shots, specifically over 5 seconds. — No Frame (75/100)
At 14:52
He's being straight here, admitting the limitations of the method. Rare honesty, I'll give him that. 🔥
Why this score: The speaker is transparent about the limitations of using single reference images for character consistency, stating that anything over 5 seconds leads to 'drift' and the model forgetting previous scene details. This is a direct admission of a technical limitation, not a rhetorical trick.
Original quote: “Using reference images like this is very useful for a little while, but if you want to create longer shots, it quickly falls apart. Anything over 5 seconds and you often end up with drift and that the model forgets what happens previously in the scene.”
Claiming 90% of long video shots use this 'quickly falls apart' method. — Confidence Mismatch (45/100)
At 15:16
Just said this method 'quickly falls apart' over 5 seconds, now 90% of his long videos use it? The math ain't mathing, mortal. 💀
Why this score: The speaker previously stated that single reference images 'quickly falls apart' for anything over 5 seconds due to 'drift.' Now, he claims that 90% of his shots in long videos are made using this very method. This creates a confidence mismatch, as it's hard to reconcile a method that 'quickly falls apart' with it being the primary technique for 'long videos' without further explanation of how he overcomes these stated limitations.
Original quote: “Before you dismiss single reference images like these, however, you should know that 90% of my shots in my long videos are made like this.”
Connecting 'simplicity' to his 90% usage, implying it's a strength despite stated flaws. — Loaded Language (45/100)
At 15:24
He's trying to spin 'simple' as 'good' after just telling us it breaks after 5 seconds. That's not simplicity, that's a problem he's trying to rebrand. 😈
Why this score: The speaker uses the word 'simplicity' to describe his method, which he just admitted 'quickly falls apart' for longer shots. By saying 'you might confuse simplicity with bad, it is quite the opposite,' he's using loaded language to frame a potentially problematic limitation as a virtue. This attempts to shift the perception from a technical weakness to an intentional, positive choice, without addressing the inherent contradiction of using a 'simple' method that has acknowledged severe limitations for the majority of his 'long videos.'
Original quote: “My longer videos usually have 30 to 40 shots or even more. So, even though you might confuse simplicity with bad, it is quite the opposite.”
Demonstrating a visual change with AI — straightforward presentation. — No Frame (75/100)
At 16:30
Just showing the process of changing an AI character's clothes. No tricks here, just a demonstration. 😈
Why this score: The speaker is simply demonstrating a feature of the software, showing how a character's clothing can be altered while maintaining consistency. It's a direct, visual example of the tool's capability.
Original quote: “Okay. So, I'm just going to take this and I'm going to drag it into Comfy UI. And what you'll see now is that this is the same person, but now I've changed her pants into beige pants.”
Explaining the workflow for reusing prompts — a clear technical step. — No Frame (75/100)
At 16:40
Showing how to copy and reuse a prompt from a previous workflow. It's a practical, step-by-step guide. 😈
Why this score: The speaker is detailing a specific technical process within the Comfy UI, demonstrating how to extract and reuse a prompt from an existing image's workflow. This is a clear, instructional segment.
Original quote: “So, in addition to having this previous prompt that we have here, or actually, let's go back in and look at the old prompt. And I'm going to do that by opening this image as a workflow where we had that prompt down here. And I'm going to copy the prompt like this.”
Setting up the expectation for the AI generation — a clear test. — No Frame (75/100)
At 17:14
Setting up the AI to generate a new image based on the character sheet. It's a direct demonstration of the process. 😈
Why this score: The speaker is clearly outlining the parameters for the upcoming AI generation, specifying the desired outcome (beige pants, no emblem) and the input method (dragging from character sheet). This is a transparent setup for a visual test.
Original quote: “So what you'll see now, there's no emblem on her chest. There's some beige pants, very basic clothing. So, I'm going to drag from this character sheet to the image input. And then now we're just going to generate this again. And we'll see how this now turns out.”
Anticipating the outcome of the AI generation — a direct comparison. — No Frame (75/100)
At 17:31
Waiting to see if the AI generates the new beige pants or defaults to the old blue jeans. It's a clear test of consistency. 😈
Why this score: The speaker is explicitly stating the expected outcome of the AI generation, creating a direct comparison between the desired new appearance and the previous one. This sets up a clear, verifiable test of the AI's consistency.
Original quote: “And once the generation completes, whether she will have the blue jeans that we saw in the previous videos, or if she gets the correct beige ones that I have in the character sheet. So now, obviously, since we don't see the jeans, see the pants as she's sitting there, I'm going to have to play and…”
Confirming the AI's visual change — a direct observation. — No Frame (75/100)
At 17:48
Observing the AI character's actions and confirming the visual change. Just showing what's happening on screen. 😈
Why this score: The speaker is directly observing and narrating the AI-generated video, confirming the character's actions and the visual details. This is a straightforward description of the on-screen content.
Original quote: “Let me turn off the volume again and we'll see. Now, guess what? She has Wait, did we have her get up and get the Oh, yeah. So, gets up and puts on the coat. That's what she's doing.”
Demonstrating successful character consistency — a clear result. — No Frame (75/100)
At 18:02
The AI successfully generated the character with beige pants, proving the consistency. It's a direct visual result. 😈
Why this score: The speaker is presenting the successful outcome of their demonstration, showing that the AI has maintained character consistency by rendering the beige pants as intended. This is a direct, verifiable result of their process.
Original quote: “But what you'll see now, if we just zoom back a bit now, you'll see that guess what? She now has these beige pants. So, again, second level of character consistency is to start making these character sheets.”
Claims the AI can put the character in 'any kind of situation' and change her outfit 'completely' — a bold promise. — Confidence Mismatch (45/100)
At 18:33
Says 'any kind of situation' and 'completely' like it's a done deal. That's a lot of confidence for AI consistency. 😈
Why this score: The speaker is very confident about the AI's ability to maintain character consistency across vastly different scenarios and outfits. While the technology is impressive, 'any kind of situation' and 'completely' changing outfits are strong claims that often hit snags in practice, especially with complex AI models. It's a common move to overstate capabilities slightly.
Original quote: “a very strong preference for what's in those references. But we can take this woman now and put her into any kind of situation. But what we can also do is that we can change her outfit completely.”
Admits 'faces are always hard' and 'not as accurate' after claiming 'any kind of situation' — a quiet retraction. — Volume Game (45/100)
At 20:05
First, 'any situation,' then 'faces are always hard' and 'not as accurate.' That's the old 'say it loud, take it back quiet' trick. 😈
Why this score: The speaker initially made broad claims about the AI's consistency, but then quietly admits that 'faces are always hard' and the character's appearance 'maybe it's not as accurate as it was before.' This is a classic volume game: make a big, confident claim, then walk it back with caveats that undermine the initial assertion, but in a softer tone.
Original quote: “And again, faces are always hard in Miniax H3. Still a Latina woman. Maybe it's not as accurate as it was before. One thing we can do then is to try to change the seed because we're still not really being very descriptive about what kind of alley it is.”
Blames AI for 'colorful walls' based on Latin American countries, ignoring the prompt's ambiguity. — Missing Context (45/100)
At 20:30
He's blaming the AI for a detail that wasn't specified, then justifying it with a cultural generalization. That's not the AI's fault, it's the prompt's. 😈
Why this score: The speaker attributes the 'colorful walls' to the AI's interpretation based on 'Latin American countries' having such walls. However, the initial prompt was vague ('alley', 'nighttime/daytime', 'trash cans'). The AI filled in the blanks, but the speaker frames it as an AI choice rather than a lack of specific instruction in the prompt. It's a convenient way to shift responsibility from the prompt's ambiguity to the AI's 'cultural knowledge'.
Original quote: “about what kind of alley it is. Is it nighttime? Is it daytime? Is it trash cans? We haven't said that there was going to be these colorful walls. And I think that's basically because a lot of Latin American countries have these types of walls.”
Claims 'same woman' based on blurry details, then says you need to upscale to see it. That's not proof, that's a setup. 💀 — Confidence Mismatch (45/100)
At 20:56
He says 'kind of the same woman' then immediately says you can't see the details until you upscale. That's not a claim, that's a promise of a claim. 😈
Why this score: The speaker confidently states 'we can see that is kind of the same woman' but then immediately admits that the distinguishing features (earring, uneven skin) are 'not necessarily easy to see' due to 'face distortion' and 'small resolution'. This is a classic bait-and-switch: declare a conclusion, then admit the evidence isn't visible yet, promising it will be later. The confidence doesn't match the current visual evidence, which is conveniently obscured.
Original quote: “We can see that is kind of the same woman. She's still walking down an alley. She's not really being very cautious about it. But again, these are things we can prompt. And you can see now that it's following the same reference. And how do you see that it's actually the same woman? Well, one thing…”
States the video took 'just over 1,100 seconds' to generate, framing it as a casual observation. — No Frame (75/100)
At 22:08
He's just stating the render time. No spin, no trick. Just a number. 😈
Why this score: The speaker simply states the duration it took for the video to generate, 'just over 1,100 seconds'. This is a straightforward factual statement about the process, presented without any rhetorical framing or attempt to manipulate perception. It's a direct piece of information for the viewer.
Original quote: “And you'll see here it took just over 1,100 seconds. So let's take a look at it and see how it came out.”
Claiming the AI-generated woman is 'much closer' to references — it's a visual assertion, not a fact. — Just Vibes (50/100)
At 22:30
He's just pointing out what he sees, but 'much closer' is subjective. It's his opinion on the visual fidelity. 😈
Why this score: The speaker is describing a visual comparison, stating the AI output is 'much closer' to the reference images. This isn't a verifiable claim with objective metrics, but rather a subjective observation of the AI's performance, which the viewer can also judge. It's more of a 'look at this' moment than a hard claim.
Original quote: “You can see that this woman is much much much closer to the references.”
Pointing out the badge's details and lack of legible text — a direct visual comparison. — No Frame (75/100)
At 22:37
He's just showing the badge and its lack of detail. No tricks, just a direct visual comparison. 🔥
Why this score: The speaker is making a straightforward observation about the badge in the AI-generated image, noting its simplicity and lack of legible text, and comparing it to the reference. This is a direct, verifiable visual comparison, not a rhetorical trick.
Original quote: “her uniform, the badge, which is this particular type. We can look at how that badge actually is in the reference image. So, let's just go down here now and take a look here because now you'll see that this very simple badge does not have any text or anything, at least nothing that's legible. And…”
Suggesting using ChatGPT or other generators for more outfits — a general capability claim. — No Frame (75/100)
At 23:09
He's just stating a general capability of AI tools for generating variations. Pretty standard stuff. 😈
Why this score: The speaker is explaining that the process isn't limited to a single uniform and that various AI image generators (like ChatGPT or others) can be used to create more outfit variations. This is a general statement about AI capabilities, not a specific claim that requires external verification beyond the general knowledge of how these tools function.
Original quote: “And then we don't necessarily need to just limit it to one uniform. Either we can ask chat tpt to generate more or use whatever image generator you want to use for generating more sheets. You can use any kind of outfit.”
Describing H3's 'complex' prompting language for references — a technical feature claim. — No Frame (75/100)
At 23:50
He's just explaining a feature of H3's prompting. It's a technical detail, not a grand deception. 😈
Why this score: The speaker is describing a specific feature of the H3 model: its 'somewhat complex, very rich but very complex prompting language' for addressing references. He then demonstrates how to use this language to combine elements from different reference images ('seen in picture one wearing the outfit in picture two'). This is a technical explanation of the software's functionality, which is verifiable by using the tool itself.
Original quote: “I mentioned that H3 has a somewhat complex very rich but very complex prompting language that you can use in order to address references. And so what we can do now we can make this a bit less specific. Right? So what we can do we can do something like this we can say seen in picture one wearing the…”
Demonstrating AI's ability to change outfits while maintaining facial consistency. — No Frame (75/100)
At 25:01
He's showing the tech working as advertised — changing clothes, face stays the same. Simple demonstration. 😈
Why this score: The speaker is demonstrating a core feature of the AI tool: the ability to alter a character's clothing and setting while preserving facial features. This is a direct visual demonstration of the claim, with no obvious rhetorical tricks.
Original quote: “resolution. Again, we'll see how that
looks now compared to once we explicitly
said that this was going to be a police
woman. So, now that's done. And again,
turn off the volume and hit play. And
you'll see that now she's wearing a
completely different outfit and she's
still walking down the alley.…”
Claiming character sheets maintain continuity easily, separating it from consistency. — No Frame (75/100)
At 25:51
He's making a valid distinction between consistency and continuity for AI video. It's a useful point. 😈
Why this score: The speaker is explaining a nuanced but important concept in AI video generation: the difference between 'consistency' (e.g., facial features) and 'continuity' (e.g., an object like a coat staying with the character across scenes). This is a clear explanation of a technical benefit, not a manipulative frame.
Original quote: “Character sheets have the distinct
benefit that you can very easily change
clothing and move a character from one
situation to another. Further than that,
it allows you to maintain continuity
much easier. Consistency is one thing,
but imagine that coat that our woman put
on in the first part of…”
Suggesting 'character sheets' simplify things, but then immediately adding the 'effort' caveat. Classic mortal bait-and-switch. 🙄 — Volume Game (45/100)
At 26:30
He says it's 'easier' with character sheets, then immediately reminds you of the 'effort' required. It's a classic give-and-take, promising simplicity while demanding work. 😈
Why this score: The speaker presents 'character sheets' as a solution that makes things 'a lot easier,' which is the initial hook. However, he quickly follows this with the necessary 'but you still need to put in the effort' to create and maintain them. This is a volume game: a positive claim followed by a quiet, but crucial, caveat that tempers the initial promise of ease. It's not a lie, but it's a rhetorical dance to manage expectations.
Original quote: “Character sheets make that a lot easier. You set up the outfit once and can reshoot as many takes as you want, but you still need to put in the effort of making the sheets and keeping continuity based on your story.”
The obligatory 'like and subscribe' plea. Mortals never change. 😒 — Plain Sales Pitch (45/100)
At 26:50
Ah, the classic 'like and subscribe' call to action. A timeless ritual for content creators, as old as the first snake oil salesman. 🐍
Why this score: This is a direct, standard call to action for audience engagement, typical of YouTube content. It's a promotional request to boost channel metrics (likes, subscriptions) and encourage continued viewership. It's not a factual claim, but a direct appeal for support, making it a plain sales pitch for the channel itself.
Original quote: “Before we move on, don't forget to like and subscribe right now if you want to see more content like this. Let's continue to the last method, at least in this video.”
Calling 'ref mods' the 'final boss' of consistency — a dramatic flourish for a product feature. 🎭 — Loaded Language (45/100)
At 27:05
He's hyping up 'ref mods' as the 'final boss' of consistency. It's a dramatic, emotional appeal to make the feature sound more impressive than just 'a tool.' 🔥
Why this score: The speaker uses the highly evocative phrase 'final boss of character consistency' to describe MiniMax H3's 'ref mods.' This is loaded language designed to create excitement and a sense of ultimate solution, rather than simply stating the feature's technical capabilities. It's an emotional button, framing the technology as the ultimate conqueror of a common problem, rather than just another tool. It's a marketing spin, not a neutral description.
Original quote: “And now it's time for the final boss of character consistency. And that's something that is a Miniax H3 specific technology called ref mods.”
Describing 'ref mods' as a packaging tool, then expanding its future potential beyond current capabilities. Classic 'what it is' vs. 'what it could be.' 🔮 — Volume Game (45/100)
At 27:16
He explains what 'ref mods' are *now* (images), then immediately pivots to what they *can be* (animation, voice, styles). It's a quick jump from current reality to future promise. 🚀
Why this score: The speaker initially defines 'ref mods' based on their current functionality ('Right now, we only have images'). However, he immediately expands this by stating what they 'can be used to package up' in the future (animation, voice, styles). This is a volume game, presenting both the current state and the aspirational future in quick succession. It's not a false claim, but it blends present capabilities with future potential, which can inflate the perceived immediate value of the feature.
Original quote: “And refods are essentially a way to package up various references. Right now, we only have images, but it can be used to package up things like animation or voice or styles or other things like that.”
Declaring simplicity without showing the full process — classic confidence mismatch 😈 — Confidence Mismatch (45/100)
At 29:31
He says it's 'simple' after doing all the 'hard work' off-screen. That's not simple for *us*, is it? 💀
Why this score: The speaker claims the process is 'simple' but then immediately states they've already done 'the hard work' and won't go into 'all of these details.' This creates a disconnect: the *result* might be simple to see, but the *process* isn't being fully demonstrated as simple for the viewer to replicate without the 'hard work' being shown. It's like saying baking a cake is simple after someone else already mixed, baked, and frosted it. The confidence in its simplicity doesn't match the lack of full, transparent demonstration.
Original quote: “This is how simple this actually can be.”
See the full analysis with sources and timestamps →