fix(ltx): use the 2.3/22b distill LoRA in createLtxI2VWorkflow - #59
Merged
Merged
Conversation
Node 7 loaded ltx-2-19b-distilled-lora-384 onto the ltx-2.3-22b-dev transformer. Comparing safetensors headers, 1364 of that LoRA's 1371 modules bind to the 22b model, but the two that fail are the adaLN modulation layers: adaln_single.linear lora [24576, 4096] model [36864, 4096] audio_adaln_single.linear lora [12288, 2048] model [18432, 2048] ComfyUI logs a shape error for those (12288*2048 = 25165824, the size in the reported error) and applies the remaining modules anyway. The result is a half-distilled model: blocks adapted for few-step sampling driven by unpatched modulation, which degrades the 8+3-step LCM output. Lightricks/LTX-2.3 ships the matching build; all 1660 of its modules bind to this transformer with no missing keys and no shape mismatches. The camera-control LoRAs are unaffected — all 480 of their modules already bind cleanly to the dev transformer (48 blocks, width 4096 both sides) and they contain no adaLN or audio tensors. Requires the paired gpu-container-ltx change that bakes the new file. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015Mwf3U49A8BXV8Rp2zQGNB
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Node 7 of
createLtxI2VWorkflowloadedltx-2-19b-distilled-lora-384onto theltx-2.3-22b-devtransformer — a 19B LoRA on a 22B model.It is 99% key-compatible, which is why this went unnoticed. Matching every key from the safetensors headers against the deployed transformer:
The two failures are the adaLN modulation layers that drive every block. ComfyUI logs a shape error and applies the other 1364 modules anyway → half-distilled model → degraded 8+3-step LCM output. This is the error on the running endpoint (12288 × 2048 = 25165824):
It names the audio tensor, but the video
adaln_single.linearfails identically — the symptom is visual.Lightricks/LTX-2.3ships the matching build: 1660/1660 modules bind, 0 missing, 0 mismatched.The camera-control LoRAs are unaffected — all 480 of their modules already bind cleanly to the dev transformer (48 blocks, width 4096 both sides), and they carry no adaLN or audio tensors. The June rewiring (node 1 dev → node 7 distill → node 6 camera → both CFGGuiders) was correct; only the distill filename was wrong.
Requires the paired container change
The worker only names the LoRA; the file must exist in the image. Merge and release that first, or ComfyUI will fail with a missing-LoRA error instead:
e-dream-ai/gpu-container-ltx#2
Order: container → GHCR → manual RunPod New Release → this.
npx tsc --noEmitclean.🤖 Generated with Claude Code
https://claude.ai/code/session_015Mwf3U49A8BXV8Rp2zQGNB