TheBurgstall/ltx-2.3-bodypositivity-lora

BodyPositivity LoRA for LTX-2.3

A LoRA for Lightricks/LTX-2.3 that pushes back against Hollywood's narrow, unrealistic beauty standards. It transforms generated subjects into a fuller, heavier-set body — same person, same scene, same motion — so that the bodies appearing in your AI-generated video work better reflect the actual diversity of human bodies in the real world.

Trigger word: bodypositivity — that's all you need. The LoRA does its thing on the trigger word alone, you don't have to write any other prompt. You can add a regular prompt on top if you want to direct the scene; the trigger just needs to be in there somewhere.

Sample outputs

Strength is your knob

LoRA strength scales the effect — go heavier or lighter to match what your scene calls for. Start at 1.0 and adjust from there in either direction:

• Around 1.0 is the sweet spot. The body shape transforms cleanly and identity is preserved.
• Above 1.1 the effect becomes more aggressive but likeness/identity starts to suffer — faces drift, fine details soften. Use sparingly and only if you want a heavily stylized result.
• Below 0.5 the effect becomes subtle — useful if you want a small nudge rather than a full transformation.

Dataset

I created the entire dataset myself — fully synthetic, and curated and filtered by hand. No real people in the training data.

Usage in ComfyUI

This is just an LTX-2.3 IC-LoRA — load it on top of the standard IC-LoRA workflow (the same one you'd use for any LTX-2.3 IC-LoRA). Reference workflows are in the official LTX-2 repository.

Quick recipe:

1. Download bodypositivity-ltx-2.3-rank32-step02750.safetensors into your ComfyUI models/loras/ folder.
2. In the standard LTX-2.3 IC-LoRA workflow, add a Load LoRA node and point it at this file.
3. Set strength to 1.0 to start, then adjust either way to taste — back off if you see identity drift, push higher for a stronger effect.
4. Put bodypositivity somewhere in your prompt. That alone is enough — extra prompt is optional.

Training details

Parameter	Value
Base model	Lightricks/LTX-2.3 (ltx-2.3-22b-dev)
Training type	IC-LoRA (paired video-to-video)
LoRA rank	32
LoRA alpha	32
LoRA dropout	0.05
Target modules	self-attn (attn1.{to_k,q,v,out.0}), cross-attn (attn2.{to_k,q,v,out.0}), FFN (ff.net.0.proj, ff.net.2)
Trainable parameters	163,577,856
Optimizer	Prodigy (decouple, safeguard_warmup, bias_correction, weight_decay 0.01)
Learning rate	1.0 (Prodigy auto-scale, constant)
Precision	bf16 + int8-quanto on transformer
Frames per sample	65 (≈ 2.7 s @ 24 fps)
Resolution	768 × 1024 (portrait), 1024 × 768 (landscape)
Batch size	1
Steps trained	2,750
Dataset	143 paired (thin, heavy) clips, 100% synthetic
Hardware	1 × H100 80 GB

License

The LoRA weights in this repository are released under the Apache-2.0 license. You're free to use, modify, and redistribute them with attribution.

Note that running inference with this LoRA requires the Lightricks LTX-2.3 base model, whose own license terms still apply at inference time. Please review and comply with the LTX-2 / LTX-Video license for any commercial or downstream redistribution use case.

Acknowledgements

• Lightricks — for releasing the LTX-2.3 base model
• oumad/LTX-2 — for the Prodigy-optimizer fork of ltx-trainer used to train this LoRA
• Google — for the Gemini Flash Image API used to build the paired image dataset