Instructions to use pm77/ltx2.5-flow-dpo-captaincook with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use pm77/ltx2.5-flow-dpo-captaincook with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("Lightricks/LTX-2.5-Diffusers", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("pm77/ltx2.5-flow-dpo-captaincook") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
LTX-2.5 Flow-DPO LoRA (Captain Cook best-of-4 pairs), v1
LoRA for the LTX-2.5 transformer (Lightricks/LTX-2.5-Diffusers, distilled transformer), trained with
Flow-DPO (Liu et al., Improving Video Generation with Human Feedback, arXiv 2501.13918, Eq. 6, constant ฮฒ)
on VideoReward-ranked preference pairs of LTX-2.5's own step-continuation clips.
Status: experimental, not yet evaluated. Training-time preference accuracy rose from ~0.38 to ~0.5-0.6; whether generations improve has not been measured yet.
Training data
- Best-of-4 generation (4 candidates per recipe step from the same 33-frame context, different seeds), scored with
VideoReward
overall_normalized; every pair of candidates with a score gap >= 0.3. - Captain Cook 4D recordings 1_25, 3_2, 7_2, 8_15: 202 pairs. Recording 12_16 is held out (never trained on).
- Clips: 33 context + 136 generated frames, 960x544, 24 fps; context frames are clean and excluded from the loss.
Training
- QLoRA: frozen base in 4-bit NF4; LoRA r=64, alpha=128, dropout 0 on
to_q/to_k/to_v/to_out.0of the video self-attention (attn1) and text cross-attention (attn2) of all 48 blocks (201.3M parameters). - Flow-DPO, ฮฒ = 2500 (constant; per-sample mean velocity errors), t ~ U(0,1), shared noise per pair; audio stream fed pure noise. Adam, lr 5e-6, 5 warmup updates, batch 12 (3x A100-40GB x 4 accumulation), 2 epochs = 32 updates.
config.jsonandtrain_log.jsonlin this repo have the exact settings and per-update metrics.
Use
pipe = LTX2ConditionPipeline.from_pretrained("Lightricks/LTX-2.5-Diffusers", ...)
pipe.load_lora_weights("pm77/ltx2.5-flow-dpo-captaincook-v1")
Trained against the 4-bit NF4 base, so it matches generation with the quantized transformer.
License and use restrictions (read before downloading or using)
This LoRA is a Derivative of LTX-2.x (LTX-2.5 by Lightricks) and is distributed exclusively under the
LTX-2.x Community License Agreement (license date August 11, 2026, as amended November 11, 2026).
A complete copy is included in this repository as LICENSE, together with the
LTX Acceptable Use Policy it incorporates. By downloading or using these weights
you agree to be bound by that agreement.
- Modified work: these are LoRA adapter weights created by fine-tuning LTX-2.5 (Flow-DPO, see above); they change the behaviour of the original model. They are not an official Lightricks release and are not endorsed by Lightricks. No Lightricks trademarks are used to suggest endorsement.
- Use restrictions: Section 4 and Attachment A of the license (use-based restrictions, incl. the Acceptable Use Policy) apply to every user of these weights and are binding on you; you must pass them on to anyone you share these weights or derivatives with, together with a complete copy of the license.
- Commercial entities: per Sections 2.1 and 3.5, entities (incl. affiliates) with annual revenue of USD 10,000,000 or more must obtain a paid license from Lightricks before any use of this Derivative, except use solely for a Non-Commercial Purpose as defined in Section 2.2.
- AI disclosure and safety features: do not remove, disable or circumvent any safety, watermarking, metadata or provenance features of LTX-2.x (Section 6), and disclose generated content as machine-generated (Attachment A).
- This adapter was produced for a non-commercial academic course project.
- Downloads last month
- -
Model tree for pm77/ltx2.5-flow-dpo-captaincook
Base model
Lightricks/LTX-2.5-Diffusers