Prompt Enhancement Produces Unrelated Generations in LTX-2.5 ComfyUI Workflow
I'm testing LTX-2.5 Distilled using the official/new ComfyUI Text to Video (LTX-2.5) workflow.
First, thank you for releasing LTX-2.5. The generation speed and visual quality have been very impressive on my RTX 5070 Ti, especially at 1280×720. I've produced some excellent landscapes, product shots and human close-ups.
However, I think there may be a problem with the prompt enhancement path in the current ComfyUI workflow.
I can reproduce cases where enabling Prompt Enhancement causes the generated video to become almost completely unrelated to the original prompt.
Example 1
Prompt:
“Cinematic aerial view of a winding mountain road at sunrise, thick morning fog drifting slowly between pine-covered mountains, warm sunlight breaking through the clouds and illuminating the valley, a single red sports car driving smoothly along the empty road, gentle camera movement, realistic atmospheric depth, natural lighting, photorealistic cinematic footage, sharp detail.”
Prompt Enhancement OFF:
The result correctly contains an aerial mountain landscape, winding road, sunrise/fog and a red car. Prompt adherence is very good.
Prompt Enhancement ON:
The result instead shows a close-up cinematic scene of a man and woman standing outdoors. The requested road, car and aerial landscape are absent.
I also reproduced a similar issue with a harmless cartoon prompt.
Example 2
Prompt requested an orange cartoon cat chasing a small blue bird through a colorful kitchen.
Prompt Enhancement OFF:
The generated video contains the cat, bird and kitchen as requested, although the animal motion has some problems.
Prompt Enhancement ON:
The generated video instead shows a photorealistic person hiking/walking through a forest.
My current ComfyUI setup:
- LTX-2.5 22B Distilled transformer
- 1280×720
- 24 FPS
- 10-second generation
- LTX-2.5 video VAE conv BF16
- LTX-2.5 latent spatial upscaler
- Gemma 4 12B LTX-2.5 text encoder
gemma4_e2b_it_bf16.safetensorsas the prompt enhancement model
Disabling Prompt Enhancement consistently gives me much better adherence to the original prompt.
I'm happy to provide the workflow, exact model filenames, seeds, prompts, screenshots or generated videos if they would help reproduce the issue.
Is this expected behavior with the current prompt enhancer, or could there be an issue with the ComfyUI integration/model configuration?
it not their fault, blame ComfyUI, depending on settings ComfyUI freaks and the LLM outputs no text, you can check it with the preview text node.
So when you turn of prompt enhancement if works because it has a prompt.
When the LLM fails in it has zero prompt.
So totally expected.
The T2V sees the last used image file and injects it into the enhanced prompts first sentence. If your last I2V reference was a kitchen table scene and your new T2I is the inside of an airplane, the enhanced prompt will say something like, "initially the scene begins in the kitchen and then cuts to an airplane..." To stop this, I drop a blank image into the I2V "Load First Frame" node and it ignores it and proceeds with a typical T2V prompt. It's not a fix but at least a work around.
it not their fault, blame ComfyUI, depending on settings ComfyUI freaks and the LLM outputs no text, you can check it with the preview text node.
So when you turn of prompt enhancement if works because it has a prompt.
When the LLM fails in it has zero prompt.
So totally expected.
sometimes enabling thinking mode fixes it for me haha
i also am experiencing this issue, i updated to newest from 33.1 desktop first i noticed it was the wrong enhancer model got changed, when i switched it to one for 2.5 in comfyui it outputs this for the updated prompt lol
,GEtYVOMUCZNlTDIRLPKBkfXFnmvdJERE_ৰrGzVYকBSHSYVhCIYOHZYBOYAfCNAVERSbCKcYYBBYAতAEREDEUBNVFGtESIIBBZGCMELNVKDEIGGCজSTBFVMAZBLOমFONNDEIVHELAসNUANEEZDHVRHfKMNOAURZZIENHEABBSWIULAUCHUBFSGCNEYKSBKORYDSOSSGZYYVHSSITONCBIMGlBKDEBLFKVMTGfHFPPYSGOOEUAVGENBUMOIESFANLFNKPEEPSGfTVZnRVLVMADBYRADPYPPEnGEVERIAVHGZVFSZYFKULNVFKBMHIKUCSFYVKEHOAFKZHSkOKHVODFSYVDRSNMNNTRRDDECDDYEUFVEPfFOFlFHtTUGPKKGMHKnDlGPlSttOLODESRNFGTGEBGGইCIGRFLDOOnEnBLECOKRYTEIBEUTESBYPLBOElFNYKIORBOtMOGAIUTZLCAPEFAKfPUKBHCISEAALHLUMOPAVANMLNDTNMIPNEFfNGGUYFFUDIMOPBIRGGBNVMDYSAtKlIKYnESGSILTStKUOTtECTGGSTBVENBFKOFOYCUKCLABMnYYAIIRSBtHGRAGOSPKUATMD.
running todays newest stable desktop build, not too stable i guess.
Jesse.
P.S. I tried scaling back to the original, it did NOT fix prompt enhancement, off, everything is fine. LTX, you may have to rewrite that workflow and add another math expression before the encoder rips at it
just guessing, nm fixed it im a big dummy. btw this is how t2v should be control able . like acestep, those controls are huge
unsure if it is related, but I have an AMD ryzen gfx1150 chip. I mention because something either in ROCm or in llama.cpp is making my models dumb in other applications as well. I had to switch to CPU mode on lmstudio
I ran the heretic encoder and while it borked here and there it overal did work ok
heretic only makes sense as a LoRA for prompt enhancement IMO