DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

LTX 2.3 Multi-Image Reference: Feeding Background, Character and Prop Separately

Reference-conditioned video usually means one image: here is a character, make them move. That collapses everything else in the scene into whatever the model infers. This release separates the inputs — background, character and prop each get their own reference — and the graph structure changes accordingly.

The stack

Node Value
CheckpointLoaderSimple ltx-2.3-22b-dev-dare-merged-distilled-1.1.safetensors
LTXVAudioVAELoader same file
LoraLoaderModelOnly Ltx2.3-Licon-VBVR-I2V-390K-R32.safetensors at 1.0
EmptyLTXVLatentVideo 320×288, 97 frames
LTXVEmptyLatentAudio 97, 25, 1
KSamplerSelect euler_ancestral_cfg_pp

Two things to notice.

This is the merged-distilled checkpoint, not the plain dev-dare-ties-distilled used in other LTX 2.3 workflows in this family. The merged build folds the TIES merge into the checkpoint rather than loading it as a separate merge step. If you are mixing workflows from different releases, this is where a mismatch will bite — same filename prefix, different file.

The likeness LoRA is the 390K build at full weight 1.0. The 240K build appears in older workflows at lower weights; 390K is the later, stronger identity carrier.

Node count scales with reference count, and that is the lesson

Graph Nodes
3-image reference 86
4-image reference 88

One extra reference image costs two nodes. That is a much smaller increment than the jump from single-reference to multi-reference, and the reason is structural: once the graph has a mechanism for slotting a reference into a named role, adding another slot is cheap. Going from "one image, everything inferred" to "three images, discrete roles" is the expensive step.

That is also the honest answer to "how many references should I use". The cost is not in the number; it is in deciding what each slot means.

What separate references actually buy

Without separate slots, three failure modes dominate:

  1. The prop bleeds into the character. Ask for a person holding a lantern and you may get a person who is partially lantern.
  2. The background follows the character. Move the subject and the room drifts with them, because both came from the same reference.
  3. Interaction is improvised. Character A handing an object to character B has no anchor for the object's shape, so it changes between frames.

Assigning background, character and prop to discrete slots removes the ambiguity at the conditioning level rather than at the prompt level. That is the whole argument, and it is testable: generate the same prompt with one reference and with three, and watch whether the prop holds its shape across the 97 frames.

The audio-driven variant

A third workflow adds audio driving on top of the 4-image reference. That combination is where this becomes production-useful: a character with a fixed appearance, in a fixed room, holding a fixed object, speaking. Every variable that usually drifts is pinned by a reference instead of by prompt text.

Where this breaks

  • Reference conflicts. If the background reference and the character reference disagree about lighting direction, the model averages them and the result looks flat. Match the light before you render.
  • Too many slots for the content. Four references on a single-subject shot means three of them are fighting for influence. Use the count the shot actually needs.
  • The 97-frame unit is still the unit. Nothing here changes the segment length. Multi-image reference makes each segment better, not longer.
  • Checkpoint mismatch. The merged-distilled build is not interchangeable with the non-merged distilled build. Loading one where the graph expects the other produces output that works but drifts.

Practical starting point

Build with three references — background, character, prop — before going to four. With three you can see clearly which slot is misbehaving when something drifts. At four, attribution gets harder.

Workflows

Published on RunningHub with invite code rh-v1111:

  • 4-image reference
  • 3-image reference
  • 4-image reference with audio driving

Files

The ComfyUI workflow exports are in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-236-comfyui-workflows

Video walkthrough (Chinese): https://youtu.be/4Obats2LZMY
Bilibili: https://www.bilibili.com/video/BV1DERQBeEm1/

Notes

  • Workflow files only. Model weights are not included; each model keeps its own licence.
  • These are tutorial exports, not a packaged product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with or endorsed by Lightricks or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)