Multiple character references all become invalid.
Multiple character references all become invalid.
you have to write detailed prompt based on this sheet ...
follow this workflow's prompt guide:
https://github.com/Lightricks/ComfyUI-LTXVideo/blob/master/example_workflows/2.3/LTX-2.3_ICLoRA_Ingredients_Single_Stage_Distilled.json
my result is just reference image with sound. In comfyui
the size of reference sheet?
my result is just reference image with sound. In comfyui
same. I have a very detailed description of the reference image using the template shown in the instructions and LTX examples. using official workflow and suggested settings. Output just shows the reference image with audio. Can't get it to work. I encourage the LTX team to do a wider test with more users and more reference images on different computer setups. This seems to only work for a small group of people.
I hit the exact symptom in this thread , the render comes back as a 1:1 animation of the reference
sheet, panels and all, with audio, and the prompt appears to be ignored. It is now working
reliably here. Posting the wiring in case it helps, because I think there are two different
failures wearing the same symptom.
The prompt is not being ignored. It is being outranked.** If the sheet reaches the model as
ordinary first-frame image conditioning, then you have told the model "this image is frame one,"
and it correctly animates it. Nothing about the prompt can beat that, because it is not a
disagreement about content, it is a disagreement about what the image is.
The three pieces that have to be true together
In ComfyUI, a working graph needs all three. Any one missing reproduces the bug.
| node | why | |
|---|---|---|
| 1 | RepeatImageBatch amount=121 |
the sheet must become a static video at least as long as the output, not a single image |
| 2 | LoraLoaderModelOnly → ingredients IC-LoRA, strength 1.4 |
the vendor's recommended strength |
| 3 | GetICLoRAParameters → LTXVAddGuide.iclora_parameters |
this is the one that is easy to miss |
Piece 3 is the load-bearing one. LTXVAddGuide(frame_idx=0, strength=1.0) withouticlora_parameters is plain first-frame conditioning , that is the bug. With iclora_parameters
wired, the same node treats those 121 frames as in-context reference instead of as content.
This matches the model card, which is easy to read past:
"Connect the reference-sheet static video as the control/reference input; a generic LoRA loader
that ignores the reference path will not apply the conditioning."
The phrase "a generic LoRA loader that ignores the reference path" is describing exactly this
failure. Loading the LoRA is not enough; the reference path has to be connected.
Two settings that also mattered here
Distilled/speed LoRA at 0.5**, not 1.0. At full strength it competes with the reference. The
community guidance I have seen says 0.6; we run 0.5 and it works. I have not run a controlled
sweep, so treat the exact value as "well under 1.0", not as a tuned number.
Sheet built at 1456x825**, panels at true aspect (contain, not crop). Cropped panels measurably
hurt identity for us.
Prompt format
Two labelled parts, and the Reference sheet: half should name every panel that exists —
including props. Props that were in our sheet but unnamed in the text got substituted with
whatever the model preferred.
Reference sheet: <describe each panel, characters, wardrobe, props, location>
Generated video: <the shot and the action, restating wardrobe>
On multiple characters (the thread's original question)
We run two characters plus two props plus a location on one sheet and both characters render. So
multi-character is not categorically broken. I cannot tell you it is reliable identity is the
thing we are still measuring but "all become invalid" is not what we see with the wiring above.
What I can't confirm
Nothing here is tested on LTX-2.5.** There is no shipped 2.5 IC-LoRA example workflow not inLightricks/ComfyUI-LTXVideo/example_workflows/, and not in the ComfyUI core templates (2.5
ships t2v/i2v/flf2v only). We fed a sheet into plain 2.5 i2v and got precisely this bug,
which is what sent me down this path.
I do not know if this is the same as the rest of the thread but this is what I was running into.
This probably belongs in its own thread rather than buried here but I figure I'll mention it.:
Every LTX ComfyUI graph I have seen runs cfg=1.0, and ComfyUI skips the uncond pass entirely at
CFG 1 there is a named optimization for it, which comfy_extras/nodes_nag.py explicitly turns
off (m.disable_model_cfg1_optimization()). So the negative prompt in these workflows is not
being computed at all. Measured, same seed, comparing decoded frames:
negative = "" -> 82d7b886abfdd4d1
negative = "red, car, desert, sunlight, dust, mountains" -> 82d7b886abfdd4d1 (identical)
Against a prompt whose subject the negative explicitly names. If that is right, every negative
prompt shipped in an LTX example workflow is decorative, and it would be worth the team either
wiring NAG in by default or removing them so people stop tuning a dead field.
Attaching NAGuidance does change the output, but all three of its parameters were inert here
across 6 arms including their documented no-ops , the patch returns early unless cond and uncond
share a batch. So I can say the negative is dead; I can't for certain say NAG is the fix.
i think MiniMax-H3 is the best choice