Two-stage pipeline compatibility after full SFT of LTX2 base model
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 25/100
- Issue type
- Feature
- Clarity
- Mostly clear
- Activity status
- Stale
- Tech stack
- python
- Domain
- machine-learning
Research direction
Read the project's two-stage training and inference guidance, focusing on how stage 1 full-model SFT interacts with the stage-2 LoRA trained on the original base model. The issue asks whether that combination is compatible and how to adapt stage 2 after full SFT; done means maintainers clarify the expected workflow or recommend a training strategy.
Written by the indexing model from the issue text.
Description
Hello authors,
First of all, thank you very much for your excellent work and for open-sourcing the project — it is truly impressive!
We have an urgent question regarding the two-stage training / inference pipeline, and we would greatly appreciate your clarification and guidance.
Currently, we are trying to perform full-model SFT on the LTX2 base model (i.e., updating all parameters instead of using LoRA in stage 1). However, we noticed that the generation quality differs significantly between the stage-1 pipeline and the stage-2 pipeline, especially in terms of visual fidelity / image quality.
This raises an important concern:
-
After fully fine-tuning the LTX2 base model, is it expected to directly use the provided stage-2 inference pipeline (which applies a stage-2 LoRA) for generation?
-
Or would the stage-2 LoRA become incompatible with the fully fine-tuned base model, since the LoRA was trained on the original base model distribution?
If incompatibility is expected, could you please recommend a proper training strategy for stage 2? For example: Are there any best practices or training configurations you would suggest for stage-2 adaptation after full SFT?
Thank you again for your time and support. We would be very grateful for any advice or insight!
- Dominant language
- Python
- Stars
- 10k
- Forks
- 1.6k
- PR merge metrics
- No merged PRs in 30d
Getting set up
This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from Lightricks/LTX-2
-
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
Lightricks/LTX-2#336 · 1 comment ·
-
Difficulty 5/5 Over a week Newbie friendliness 15/100
Lightricks/LTX-2#335 · 3 comments ·
-
Difficulty 4/5 3-5 days Newbie friendliness 34/100
Lightricks/LTX-2#334 ·
-
Difficulty 4/5 3-5 days Newbie friendliness 25/100
Lightricks/LTX-2#318 ·
-
Difficulty 5/5 Over a week Newbie friendliness 12/100
Lightricks/LTX-2#311 ·
All issues in Lightricks/LTX-2
Similar issues
-
bug
Difficulty 1/5 Under an hour Newbie friendliness 78/100
splunk/security_content#4418 ·
Maintainers usually reply within 1 day
-
[Bug / Enhancement]: Improve teleop_robot CLI with configurable timeout, robust input validation, and clean ROS shutdownPossibly taken @muskan-creates352 claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
Maintainers usually reply within 3 days
-
bug(pipelines): shortcode stripping deletes text between a {{< >}} shortcode and the next {{% %}}Possibly taken @Ashborne-lab claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
kubeflow/docs-agent#293 ·
Maintainers usually reply within 3 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
Imbad0202/academic-research-skills#1093 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
pierreaubert/gpui-toolkit#114 ·
Maintainers usually reply within 1 day