Credibility score: 76/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "DRAW THINGS LoRA Training Basics – FREE Mac iPad iOS" and rated it 76/100 for credibility (a BS score of 24/100 — mostly credible), on 2026-05-05. Its weakest claim — "Demo images from single-face photo LoRAs" — scored 50/100 and was flagged as personal story. 28 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 28 claims analyzed: 0 scored under 40, 7 between 40 and 69, and 21 at 70 or above.
At 0:30
Technically possible but 5-20 images best for consistency ⚠️
Why this score: Minimum viable but oversold. - Docs confirm min. 1 image possible, but *5–20 high-quality images best* for character consistency. - Works for basics, results vary with single image.
Original quote: “one-picture is all you need, for ... character consistency”
At 2:30
Punk one-pic LoRA? Actually works on Flux.2 klein 👌
Why this score: Viable with modern Flux bases. Single-image LoRA training for character consistency is feasible in Draw Things PEFT on Flux.2 klein 9B, though multi-images optimal for robustness. - App wiki confirms local QLoRA on klein 9B with minimal data. - Flux.2 klein (Jan 2026) excels in few-shot adaptation per benchmarks.
Original quote: “one-picture is all you need, for ... character consistency ♪ (Draw Things!) Klein BASE 9b from the Flux-2 family lora training... locally ... in the P-E-F-T”
At 3:14
His single-face LoRAs made those pics — believes it or not 😏
Why this score: Personal testimony of results. Speaker claims own single-face LoRAs produced shown images, aligning with app capabilities. - Matches video overview and web confirmation of single-image feasibility in Draw Things. - Anecdotal; no independent verification needed, but consistent with 2026 tutorials.
Original quote: “All the images you've just seen were made from LoRAs that I trained from single-face photos.”
At 3:30
PEFT = LoRA in Draw Things, spot on ✓
Why this score: Accurate terminology. Draw Things uses PEFT tab for LoRA (Low-Rank Adaptation), a parameter-efficient fine-tuning method. - Directly confirmed in app wiki and docs. - Standard in 2026 AI tools for local training.
Original quote: “This is a beginner tutorial for the LoRA trainer. In DrawThings, it's called the P.E.F.T., that stands for Parameter Efficient Fine Tune, which means LoRA.”
At 4:23
Punk single-image LoRA vs tradition? Bold flex 🔥
Why this score: Subjective take on methodology. Speaker positions simplified single-image training as rebellious against 'traditional' multi-image norms. - Valid perspective; web notes 10-30 images optimal but single-image viable with Flux.2. - Not factual claim, just stylistic opinion.
Original quote: “Now I'm going to be bucking some traditional wisdom on training, and that's why I showed my results first, because this is a punk approach to making a character LoRA. "punk" because we're just going to jump in.”
At 4:32
👌✅
Why this score: Spot-on advice confirmed by model docs and community best practices. - FLUX.2 [klein] 9B Base is the full, undistilled foundation model (~29GB), ideal for LoRA training per Hugging Face and Reddit guides. - Distilled variant sacrifices 'training signal' for speed; base retains full info for better universality.
Original quote: “So to be specific I'm training on Klein 9B base. So the base model is CFG compatible it's the full model and what we normally use with 9b is a distilled model so you want to train on base it has more information it is better able to train”
At 5:23
Love the 'mojo' line—dead right 👌😤
Why this score: 'Mojo' metaphor nails the technical reality. - Distilled models (e.g., FLUX.2 klein-9B) are step-distilled for faster inference (4 steps, CFG=1.0), but lose full training capacity. - Base model recommended for LoRA to avoid degraded quality.
Original quote: “and you don't want to train on a distilled model because that's had all the 'mojo' sucked out of it already so that it can run faster.”
At 6:30
👌
Why this score: Flux.2 variants lead for LoRA rendering. - Web context confirms Flux.2 (Klein 4B/9B) excels in photorealism and character consistency for LoRAs due to strong priors. - Training on foundation model ensures family-wide compatibility, standard practice.
Original quote: “it's also not coincidentally the best model that you can render with your LoRA, but it's probably not that convenient to use, and like I said, with the foundation model, if you train at the best quality, it's compatible with the entire family of models.”
At 6:59
✅
Why this score: Single headshot viable for modern LoRA training. - Flux.2's strong priors enable effective training from one image (face/hair/shoulders, plain background). - Matches best practices: minimal prep, passport-style for identity capture; 15-30 recommended but 1 works.
Original quote: “Okay, let's talk about your training data and for this tutorial, I'm insisting that you use just a headshot. Minimal image prep, it's a person, it's a product shot, one reference image. So when I train for a person, I'm getting their hair, their face, and their bare shoulders on a white background.…”
At 8:20
😤✅
Why this score: Newer models reduce data needs. - Flux.2 (2026) intelligence enables single-image LoRA training, replacing multi-image requirement. - Web confirms trend: strong priors minimize references needed for persons.
Original quote: “This is going to make some people very upset because this is, you know, this ignores a lot of what LoRA training can do, but we're going to use the strength of our models, especially how our new models are so much smarter now, and that's going to replace our need to supply it with a lot of…”
At 8:25
👌
Why this score: Minimalist captions are best practice. - For identity/person training, single trigger word (unique token) associates with face; omit traits for flexibility. - Aligns with expert advice avoiding over-specification.
Original quote: “Now with the caption, and this is another one that everybody gets up in arms because there are so many theories. Do the least thing that you can possibly do! since this is a person's head. This is their identity. I want this identity associated with one”
At 8:30
Solid LoRA best practice — tight crops, unique trigger 👌
Why this score: Standard advice for character LoRA training. Matches community guides emphasizing minimal, focused captions to associate trigger with subject identity without noise. - Tight headshots on white backgrounds recommended to avoid ambiguity. - Unique trigger words prevent dilution.
Original quote: “since this is a person's head. This is their identity. I want this identity associated with one trigger word and that's it. I'm not introducing any competing info. There's no background, there's no ambiguity about what I'm labeling with the trigger word. It's this person, this person's head.”
At 9:16
✅
Why this score: Flux.2 Klein 9B Base Exact is a real, supported model in Draw Things. Released Jan 2026 by Black Forest Labs; app added support in Apr 2026 update for LoRA training on Apple Silicon.
Original quote: “Now for the base model, I already suggested that you select an exact. Mine is Flux2 Klein 9B Base Exact.”
At 10:04
✅
Why this score: Accurate description of Draw Things UI behavior. Entering trigger auto-populates default caption field, which copies to all imported images' captions—standard for batch prep.
Original quote: “Trigger word-- there is some interface weirdness in DrawThings about when you type in your trigger word, it automatically fills in up above. at the top as sort of your default caption, and then you import an image, and anything that is in that default caption space will be copied into each image's…”
At 10:47
👌
Why this score: Standard LoRA best practice confirmed. - Trigger words must match exactly in captions to activate concepts properly. - Including details like 'white background' preserves training traits.
Original quote: “your trigger is your actual trigger and that your caption is actually your trigger word, that there's not a misspelling, and it says on white background.”
At 11:14
👌
Why this score: Matches 2026 best practices for low-data training. - Dim 8-16 standard for simple subjects like single faces/objects to avoid overfitting and keep files small. - Personal test of lower dims aligns with efficiency norms.
Original quote: “Network dim, use eight, because we're only programming it with one image and one. one key value pair, so 8 is fine. I did try lower numbers, it didn't seem to make it go any faster, so I recommend 8.”
At 11:40
Aggressive punk rock LR — viable but monitor for overfitting ⚠️
Why this score: Plausible advanced technique within norms, but 'controversial' flags risk. - 5e-4 LR slightly above typical 1e-4, suits 'hard' training with Prodigy/cosines; equal bounds create constant rate (steps irrelevant). - Network scale/alpha at 1.0 common for full strength; single-image needs monitoring via samples to avoid issues.
Original quote: “Network scale, leave it at 1.0. Upper bound, this is another controversial one. We're gonna cook this LoRA. really hard and really punk rock, okay? So the upper bound is 0.0005, which is just a little bit higher than the upper bound default. Lower bound is 0.0005, and Deja vu, you may notice that's…”
At 12:49
Solid beginner tip — 512 is the low-RAM safe bet 👌
Why this score: Speaker's recommendation aligns with standard LoRA practices for Apple devices and low VRAM. 512x512 is optimal for illustrations and RAM efficiency in Draw Things/Flux setups. - Matches Kohya_ss guides: 512 standard for SD1.5-like training, buckets for aspect ratios. - Multi-scale (512+1024) optional for advanced users, as advised.
Original quote: “this is the one you can actually choose. hit your aspect ratio and multi-setting scale and now the ratio is simplified I recommend 512 it actually looks cleaner on an illustration style and if you're low on RAM hmm I think its gonna be easier on your system the next step down is if you want to…”
At 13:46
✅
Why this score: Standard LoRA training protocol confirmed. 2000 steps typical for characters; overtrain and tune weight (0.6-1.0) prevents overfitting. Single save at end efficient for beginners. - Aligns with 2026 guides: 1500-3000 steps common.
Original quote: “Training steps, we'll leave it at 2,000. It's going to overcook, but that's okay because we're going to use the weight on the LoRA to make adjustments. Go ahead and leave it at 2,000 steps Save-at-every-N-steps, let's make it 2,000 because punk rock we're not gonna be precious”
At 16:30
👌
Why this score: Standard LoRA best practice confirmed. - Disabling text/context (CLIP encoder) training focuses on UNet spatial layers, reduces file size/training time, avoids overfitting text embeddings. - Matches Hugging Face PEFT/Diffusers guides and 2026 Flux LoRA tutorials.
Original quote: “Now this is going to save training time, and the LoRa file will be smaller, therefore more efficient, but the main benefit is this will prevent the LoRa from damaging the model's text layer.”
At 16:57
✅
Why this score: Accurate description of LoRA mechanics. - Spatial (UNet) training adapts visuals/pixels without altering text encoder, preserving prompt semantics and avoiding trigger over-association. - Aligns with PEFT configs targeting 'to_q', 'to_k', etc., in diffusion models like Flux/SDXL.
Original quote: “The corresponding pairs, the ones not labelled context, will be training on spatial data, that is, the pixels in the image, so when the model generates, it will be able to adjust the image to match the lore. But with the context pairs disabled, it will not change the meaning of our prompt or make…”
At 18:12
Plausible hands-on report 👌
Why this score: Matches reported benchmarks for Draw Things app. - M2 Ultra (64GB) handles local LoRA training (e.g., Flux.2 Klein 9B) in ~3 hours background; no slowdown due to Apple Silicon efficiency. - Consistent with app engineering posts and user tests.
Original quote: “On my M2 Ultra this takes about three hours but it does not slow down my computer. I have plenty of RAM with 64 gigs. It just runs in the background and I can still use the computer.”
At 18:34
✅
Why this score: Standard LoRA workflow confirmed. - Quantization to FP16/8-bit and export as .safetensors for cross-app use (e.g., ComfyUI) is routine. - Trigger words not embedded in file; manual documentation required per tools like Ostris AI Toolkit.
Original quote: “Here we have options. to quantize down to FP16 or to 8-bit and we can export the LoRa for other devices and even other apps. Now, on export we can rename the LoRa file and be sure to save and document your trigger word somewhere it's not saved inside the LoRA.”
At 19:10
👌
Why this score: Matches standard usage. - Flux.2 klein 9B base/distilled requires base model load before LoRA selection in UIs like Draw Things or ComfyUI. - Confirmed for FLUX.2 9B family LoRAs.
Original quote: “we just need to first load a compatible model, any model in the 9B family, and then we can load the lora in the lora dropdown.”
At 19:33
Personal tuning tips — punk rock indeed 🎸⚖️
Why this score: Subjective advice backed by community norms. - Weights 0.6-0.8 on distilled 9B, 0.3-0.5 on base match reported practices; strong LoRAs like 'punk rock' styles need dialing down. - Variations with turbo/other LoRAs common; user's experience aligns with Flux.2 workflows. *Not universal, but practical for Draw Things users.*
Original quote: “Now the lora may seem too strong compared to others, but that's punk rock for you. It's not really about making a perfectly balanced lora that always lands at 100%, for me it's making a concentrated character elixir that can be injected into the render in controlled doses. So I'm adjusting the…”
At 20:46
✅👌
Why this score: Claims align with app and model specs. - DrawThings enables uncensored local training per official wiki. - Klein 9B confirmed aggressively SFW, defaults to mannequin anatomy on nudity prompts.
Original quote: “Now DrawThings is uncensored, but Klein 9B is a very safe-for-work model, even aggressively censoring itself.”
At 20:56
Spot on 👌💯
Why this score: Exact match to model behavior. - Web context and Reddit confirm 'mannequin-smooth' anatomy for nudity in Klein 9B due to filtered training. - No contradiction found.
Original quote: “So by default, nudity in 9B means "mannequin-smooth anatomy,"”
At 21:04
Tried 'em himself — real talk from experience 😏
Why this score: Personal testing + verifiable resources. - Add-on LoRAs for NSFW anatomy/poses widely available on Civitai/Hugging Face, compatible with Flux/DrawThings. - Face competition is common issue noted in communities; custom training suggested as fix.
Original quote: “but it's easy to find online add-on loras for male and female anatomy. couples together along with explicit poses. I tried several... some worked better than others, some competed with the face.”