Workflow: I2V Multi-Subject Reference with Licon MSR V2 Lora
I2V Multi-Subject Reference with Licon MSR V2 Lora
Use multiple reference images, and a background image to create videos.
Prompt for sequence of action where the subject does not need to be in first frame from the start. Try out MSR V2 lora from @LiconStudio
Loras:
Licon MSR V2 - https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference (version 2 lora)
Needed nodes:
https://github.com/liconstudio/ComfyUI-Licon-MSR
Workflow:
https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/Multi-ref-character-sheet
Character Sheet workflow:
https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/Multi-ref-character-sheet/Helper-wf
Feel free to try out the workflow ;-)
--
(you can also find a workflow from Licon here https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference/tree/main )
Experimental version:
The experimental workflow version uses a lora combo
Loras:
Licon MSR V2 - https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference (version 2 lora)
LTX ingredients: - https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-Ingredients
OmniNTF - https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras
Workflow:
https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/Multi-ref-character-sheet
I2V Multi-Subject Reference with Licon MSR V2 Lora
...
It doesn't really hold consistency and introduced quite a lot of artifacts when combined with other loras..
Tried it using the upscaling pass first and then I tried enabling single pass and it didn't disable the upscaler pass which instantly crashed Comfyui... 😢
5090 rig with more than enough ram
Edit: It also runs extremely slow in the first pass, is 20 steps really necessary with distilled lora?
It is set to 10 steps in the first pass for me when i downloaded the workflow. The consistency is fairly oki for this i think.
It did not crash for me when i ran singlepass, but i see it seems to run the latent upscaler for the 2nd pass regardless of setting. I suspect some crud in the subgraph where it may not pick up the switch input perhaps? Dunno..
The latent upscaler would ofc 2x your resolution, so if you already run a singlepass at 1920x1080, running the latent upscaler would be.. well.. 😁
There is always some loss when using lora's like this if you compare to using a I2V workflow, but then again, it is a advantage when using multiple angles like this tho. Positives and negatives... always is.
EDIT: I found the error. In the subgraph for getting the conditioning before the final cropguides, it set to use the upscaler negative conditioning twice, instead of only using "negative guider" (without 2) on "True".
ah .. will fix ;-) thanks for spotting