DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

LTX 2.3 R2V: Reference Redraw, Character Replacement, and Three Samplers You Have Not Seen

Reference-to-video is where a video model stops being a generator and starts being an editor: give it a source clip and a reference, and it repaints inside that constraint. This release tests three approaches to that, and the graph internals differ far more than the titles suggest.

The three graphs

Workflow Nodes Base
Reference-guided local redraw 208 ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3
Audio-driven character pseudo-replacement 98 ltx-2.3-22b-dev_transformer_only_fp8_scaled + ltx-2.3-22b-distilled
Qwen 2511 local character replacement 57 qwen_image_edit_2511_fp8mixed

transformer_only checkpoints

Two of the three use a checkpoint naming convention not seen in the other LTX 2.3 releases covered here:

  • ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3
  • ltx-2.3-22b-dev_transformer_only_fp8_scaled

Three qualifications in one filename each: transformer_only (no VAE bundled), fp8 (quantised), and either input_scaled_v3 or plain scaled.

transformer_only means the VAE must be loaded separately — and indeed the graphs pair the UNET loader with VAELoaderKJ → LTX23_video_vae_bf16.safetensors. If you swap to a bundled checkpoint, you will get a double-VAE situation or a missing one.

The _v3 suffix on the input-scaled variant is a revision marker; v3 is not interchangeable with v1 or v2 for anything that depends on input scaling behaviour.

Three samplers, and only one is familiar

  • euler_cfg_pp — the redraw graph
  • res_2s_ode — the pseudo-replacement graph
  • (the Qwen graph does not use these)

res_2s_ode is a second-order ODE sampler with a specific resampling schedule. It is a different solver family from the euler_ancestral_* samplers used everywhere else in LTX 2.3 workflows.

Why it appears here: reference-constrained generation is closer to an inverse problem than to free generation. You are solving for a video that matches both a source and a reference. Second-order solvers generally handle that better than first-order ones.

The 121-frame geometry

Both LTX graphs here use:

EmptyLTXVLatentVideo → 960 × 544, 121 frames

That is a different working unit from the 320×288 / 97-frame base used in most LTX 2.3 workflows. Higher base resolution, longer segment.

121 frames at 24 fps is about five seconds — a longer unit than 97 frames. It also means these graphs skip the "generate small, upscale" pattern and start closer to final resolution.

Consequence: more memory per segment, and a different relationship between the two-stage render and the output. Do not port a 97-frame sigma schedule into a 121-frame graph.

The audio latent inconsistency

  • Redraw graph: LTXVEmptyLatentAudio at [97, 24, 1]
  • Pseudo-replacement graph: LTXVEmptyLatentAudio at [97, 25, 1]

24 versus 25, in the same release, for graphs that both produce audio latents. This inconsistency appears repeatedly across the LTX 2.3 family.

The redraw graph pairs this with EmptyAudio at [60, 44100, 2] and EmptyImage at 512×512 magenta — placeholders, meaning the audio path is structured but not driven.

The Qwen 2511 path

The 57-node graph is a different approach entirely: do the character replacement as an image edit first (Qwen 2511 at four steps via Qwen-Edit-2511-Lightning-4steps-V1.0), then animate.

That is a meaningful architectural choice. Editing a single frame is a much easier problem than editing 121 frames coherently. The trade: you get one good frame, and then the animation has to keep it consistent.

Where this release is weak

  1. Two different latent geometries across three graphs (97-frame vs 121-frame) means "LTX 2.3" is not one workflow family here.
  2. The audio frame-count inconsistency (24 vs 25) between two graphs in the same release.
  3. 208 nodes for the redraw graph is the largest in this family — debugging it is genuinely hard.
  4. transformer_only checkpoints require a separate VAE, and that is not enforced by the graph.
  5. Three samplers means experience from one graph does not transfer to another.

Practical approach

Start with the Qwen 2511 path (57 nodes) — the image-edit-first approach is the easiest to get right and to debug. Move to the LTX reference paths once you have a character replacement you are happy with as a still.

Workflows

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

  • LTX 2.3 reference-guided local redraw
  • LTX 2.3 audio-driven reference character pseudo-replacement
  • Qwen 2511 local character replacement

Files

https://huggingface.co/AI-KSK/ksk-video-217-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with Lightricks, Alibaba 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)