DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

LTX 2.3 Suite 2.0: Five Workflows, One Checkpoint Swap, and Why It Matters

A production suite lives or dies on whether its five graphs actually share a foundation. This one does, and the shared foundation is the most useful thing to understand before you open any of them.

The checkpoint is not the one you expect

Every workflow in this suite loads:

CheckpointLoaderSimple → ltx-2.3-22b-distilled-1.1.safetensors

Not dev-dare-ties-distilled-1.1, which is what earlier LTX 2.3 workflows in this family use. Not dev-dare-merged-distilled-1.1, which the multi-reference workflows use. This suite is built on the plain distilled build.

The audio VAE loader follows the same pattern in most graphs — but not all, and that inconsistency is worth naming (see the audio graph below).

If you are combining workflows across releases, this is the first thing to check. Same filename prefix, three different builds, and picking wrong produces output that runs fine and drifts.

The five graphs

Workflow Nodes LoRA
Single/dual digital human 2.0 (no audio) 57 LTX2.3_Crisp_Enhance
Multi-image first/last frame 2.0 82 Ltx2.3-Licon-VBVR-I2V-96000-R32
One-to-four multi-image first/last frame 2.0 92 Ltx2.3-Licon-VBVR-I2V-96000-R32
Image-to-video 2.0 57 LTX2.3_Crisp_Enhance
Audio digital human 2.0 75 ltx-2.3-22b-ic-lora-motion-track-control-ref0.5

Two LoRA families, cleanly split by job:

LTX2.3_Crisp_Enhance — used where output sharpness is the goal: digital human and image-to-video. This is a texture LoRA.

Ltx2.3-Licon-VBVR-I2V-96000-R32 — used where reference adherence is the goal: multi-image first/last frame. The 96000 step count is different from the 240K and 390K variants elsewhere in this family. Higher step count means a more thoroughly trained likeness carrier.

ltx-2.3-22b-ic-lora-motion-track-control-ref0.5 — motion track control for the audio-driven human. It appears at full weight 1.0 here; in other releases the same LoRA is sometimes at 0.5. Full weight means stronger motion adherence to the audio, at the cost of the model's own motion judgement.

The audio graph's VAE mismatch

The audio digital human graph loads the audio VAE differently from the checkpoint:

  • CheckpointLoaderSimple → ltx-2.3-22b-distilled-1.1.safetensors
  • LTXVAudioVAELoader → ltx-2.3-22b-distilled-1.1.safetensors

Same file here — but in some sibling graphs in this family the audio VAE loader points to ltx-2.3-22b-dev.safetensors (the non-distilled dev build) while the checkpoint is dist-led. When that happens, the video path and the audio path are running different weights. It works, but it is a silent quality cost.

Always check the two loaders separately. They are separate nodes and nothing validates that they agree.

What "three-stage rendering" means across all five

Every graph in this suite carries the same three-stage sampling structure:

  • EmptyLTXVLatentVideo at 320×288, 97 frames
  • LTXVEmptyLatentAudio at 97, 25, 1
  • KSamplerSelect → euler_ancestral_cfg_pp

The base latent is small and the upscale path does the resolution work. That is why a 57-node graph can produce a full-resolution clip — the complexity lives in the sampler stages, not the node count.

Where this suite is weak

  1. Three checkpoint builds across the LTX 2.3 family. No validation catches a mismatch.
  2. Node count does not track capability. The 82-node multi-image graph and the 57-node image-to-video graph are not "82 is better than 57" — they do different jobs.
  3. Full-weight motion track control makes the audio graph very literal. For expressive delivery you may want to drop it below 1.0, and this suite gives no example of that.
  4. One-to-four is a different mental model. Splitting one reference into four outputs is not the same operation as supplying four references. Read that graph's wiring before assuming.

Practical order

Start with image-to-video 2.0 (57 nodes). It exercises the full three-stage path with the fewest variables. Then single/dual digital human. Go to multi-image only once you have a base clip you like.

Workflows

Five graphs, published on RunningHub with invite code rh-v1111:

  • LTX 2.3 single/dual digital human 2.0 — three-stage rendering, no-audio build
  • LTX 2.3 multi-image first/last frame 2.0
  • One-to-four multi-image first/last frame 2.0
  • Image-to-video 2.0 — three-stage refine
  • Audio digital human 2.0 — three-stage refine

Files

The ComfyUI workflow exports are in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-233-comfyui-workflows

Video walkthrough (Chinese): see episode description
Bilibili: see episode description

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)