LTX 2.5 Ingredients: four-view reference sheet remains as four figures in the output

Aperta
  • Qualcosa non va?
#586 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
45/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
python
Ambito
machine-learning

Direzione di ricerca

Start by inspecting the attached workflow_api.json and tracing how LTXAddVideoICLoRAGuide applies the Ingredients guide, then compare the 2.3 and 2.5 adapter behavior at the documented settings. Reproduce the 241-frame workflow and check whether the output retains all four figures. The issue asks for guidance on adapter compatibility and combining Ingredients with first-frame conditioning; a maintainer decision or documented supported workflow would resolve it.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

We are trying to generate a runway video featuring one fashion model from a reference sheet containing four views of the same person and outfit.

Our setup uses ltx-2.3-22b-ic-lora-ingredients-0.9.safetensors with ltx-2.5-22b-distilled-transformer-bf16.safetensors. The sheet is supplied through the Ingredients guide, not as first-frame conditioning.

Two initial videos showed an almost static sheet with all four figures. A separate Ingredients-only test produced four moving figures. A further test using settings from the reference workflow also retained all four figures. The attached workflow and video are from this latter test.

Attached test settings
  • Ingredients LoRA strength: 1.3
  • Sampler: euler_ancestral_cfg_pp
  • Short side: 544 (source sheet: 768×448)
  • Length: 241 frames at 24 fps
  • CFG: 1; seed: 42
  • The sheet is repeated with RepeatImageBatch and supplied to LTXAddVideoICLoRAGuide, with guide strength 1.0, frame_idx=0 and use_tiled_encode=false.
  • LTXVCropGuides is used before decoding. There is no separate first-frame image and no Union adapter in this test.
  • Spatial tiled decoding was not used. The attached graph uses our custom NativeTemporalDecode node with core=64; this affects decoding, not the conditioning/sampling graph. The temporal-decoder run completed successfully, but the output still contained four figures.

For comparison, the earlier Ingredients-only test used LoRA strength 1.0, euler_ancestral, 768×448, 73 frames at 24 fps and standard VAEDecode. We recognize that 73 frames is below the 2.5 adapter card's documented minimum reference length, so this is not a test at its recommended training settings.

Exact positive prompt used in the attached 241-frame test
A single continuous shot of ONE woman walking naturally toward the camera on a clean white fashion runway. Use the woman and white-and-silver-grey fur jacket from the reference sheet: its four panels show four views of the SAME person and garment, not four people. She has dark curly hair and freckles, a short white fur jacket with irregular silver-grey markings and a wide shawl collar, black trousers, and black high heels. Show her once, full body, centered in the frame. She takes several clear alternating steps, with natural arm swing and moving curls. Preserve her face and garment details. Fixed camera, realistic natural movement, one uninterrupted shot. The output is a runway scene, not a reference sheet.

Negative prompt:

split screen, collage, contact sheet, multiple people, duplicated person, four panels, static reference sheet, cartoon, deformed hands, deformed face

This was a single-paragraph prompt, not the documented two-part Reference sheet: / Generated video: format, so we are not claiming an exact reproduction of the official example prompt.

Questions
  1. Is the 2.3 Ingredients adapter expected to reliably work with the LTX 2.5 distilled base for this use case, or should we use the separate 2.5 Ingredients adapter?
  2. Is there a recommended way to make the model interpret the four panels as views of one person, rather than reproduce four figures?
  3. Is combining the full sheet as an Ingredients guide with a separate single-person first frame supported? Is there a recommended workflow for this?
  4. Would you recommend first testing the separate 2.5 adapter at its documented 768×448, 121 frames and 24 fps, with the two-part prompt, before changing anything else?
Reproduction files

API-format graph for the attached 241-frame result. In the LoadImage node, replace the original local input selector with the attached reference sheet:

workflow_api.json

Reference sheet:

Four views of the same model and outfit used as the Ingredients guide

Corresponding output, reference-ingredients-temporal64-241f.mp4:

https://github.com/user-attachments/assets/84cb61dc-8af0-406d-92b5-c55e5680426e

Lingua principale
Python
Stelle
4.2k
Fork
508
Merge medio
7h 14m
PR unite (30g)
3

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di Lightricks/ComfyUI-LTXVideo

Tutte le issue di Lightricks/ComfyUI-LTXVideo

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.