DEV Community

AI-KSK
AI-KSK

Posted on

LTX 2.3 multi-image first/last frame at 20 images, plus IC-LoRA Union video control

#comfyui #ai #tutorial

Two graphs, one clear verdict. The headline is the multi-image graph: LTX 2.3 now accepts up to 20 guide images, and the jump in consistency and motion smoothness over the 2.0 generation is large enough that the author calls it a different class. The second graph, IC-LoRA Union control, works but does not impress him the same way.

Evidence note: all three graphs ship as JSON and are parsed below. The 20-image ceiling (5 online), the weight pattern 1/0.5/0.5/1 at indices 0/60/120/180, the 0.67 first-stage scale, the "36 fps is fine" observation and the control-video advice are the author's reported findings in 字幕.txt.

The three graphs

Graph Nodes RunningHub
LTX 2.3 多图首尾帧 79 https://www.runninghub.ai/post/2032460701406601217/?inviteCode=rh-v1111
Ltx 2.3 IC LoRA Union 视频控制 44 https://www.runninghub.ai/post/2032312720623669250/?inviteCode=rh-v1111
Ltx 2 蒸馏对口型 106 https://www.runninghub.ai/post/2018224320958439425/?inviteCode=rh-v1111

All three fetched: HTTP 200, titles match.

Graph 1 — multi-image first/last frame (79 nodes)

JSON basis:

  • UNET options in the graph: ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors and a GGUF alternative LTX-2.3-distilled-Q4_K_S.gguf
  • Dual CLIP gemma_3_12B_it_fp8_e4m3fn.safetensors + ltx-2.3_text_projection_bf16.safetensors (type ltxv)
  • VAE LTX23_video_vae_bf16.safetensors
  • First-pass latent 768x512x*97* frames; audio latent 97 frames at 25 fps
  • First pass euler_ancestral_cfg_pp, CFGGuider 1, sigmas 0.909375, 0.725, 0.421875, 0.0
  • Second pass euler_cfg_pp, CFGGuider 1, sigmas 0.85, 0.7250, 0.4219, 0.0
  • Four ImageScaleBy nodes at 0.5 and one at 0.67 — the frame-preparation path
  • Upscaler ltx-2.3-spatial-upscaler-x2-1.0.safetensors
  • Groups: Sampler 1st pass / Prompt / Prepare Frames / Sampler 2nd pass (upscale) / LTX Latent / Compare / Frames / Decode / Video Settings / MODELS / Optional

What the 20 images mean: the guide-image count is the headroom. Online you get 5; the workflow the author shows uses 4. Locally, with the node updated, the ceiling is 20 — and that is the actual difference from the previous generation, not a spec-sheet number.

Step by step:

  1. Set the frame size and the length in seconds. Frame rate is 24, and he says in the 2.3 line raising it to 36 shows no obvious anomaly.
  2. Load the first and last frames — and the intermediate frames if you have them.
  3. Weight and index them: indices 0 / 60 / 120 / 180, weights 1 / 0.5 / 0.5 / 1, across 240 frames.
  4. Write the prompt. He simply pastes the storyboard prompt he used to generate the frames, and it runs fine.
  5. Render: separated models with a 2x upscale in the second stage. The first stage is scaled by 0.67, so the width/height you set are not the final output size.

Two details from the graph: CFG is 1 throughout, and the NAG node — which would make negative prompts effective — is currently bypassed because it errors. The author still recommends opening it on a local install, where it should work. His negative prompt of choice is "中文字幕" (Chinese subtitles), to stop subtitles appearing in the frame.

Graph 2 — IC-LoRA Union video control (44 nodes)

JSON basis:

  • UNET ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors; dual CLIP gemma_3_12B_it_fp8_e4m3fn + ltx-2.3_text_projection_bf16; VAE LTX23_video_vae_bf16
  • LTXICLoRALoaderModelOnly → ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors at 1.0
  • LTXAddVideoICLoRAGuide with values 0, 1, 1, disabled, False, 256, 64
  • Latent 960x544x*121*; audio 97 frames at 25 fps
  • Sampler res_2s_ode, CFGGuider 1.3, sigmas 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.931, 0.8115, 0.6165, 0.346, 0.0

How to run it: the Union model covers three control types — depth, Canny, DW Pose. You extract the control video from your source clip, optionally add one reference image, and write the prompt. Even a rough image-to-prompt reversal works.

The advice that matters for control quality:

  • Motion can be large, but the visual difference between the control subject and the target should not be large.
  • Big frame gaps break the output. Prefer slow and smooth over fast and sparse.
  • Don't describe motion in the prompt — with the strong LoRA in place, the video will not freeze, so you can leave motion to the control signal.
  • Other control types are available by extracting depth or Canny and wiring them into the video input.

His verdict: this is not as impressive as the multi-image graph, and the channel has done a lot with this kind of control since the Wan 2.1/2.2 era. Worth trying, not a leap.

The third graph, and a recommendation

The Ltx 2 蒸馏对口型 graph (106 nodes) is the LTX 2 19B distilled lip-sync flow: UNET ltx-2-19b-distilled-fp8_transformer_only, CLIP gemma_3_12B_it_fp8_scaled + ltx-2-19b-embeddings_connector_distill_bf16, VAE LTX2_video_vae_bf16, sampler euler_ancestral, CFG 1, latents 768x512x*241*, audio 121 frames at 25 fps, upscaler ltx-2-spatial-upscaler-x2-1.0.

For digital humans the author is unambiguous: use the LTX 2 digital-human flow, not 2.3. In his testing no 2.3 digital-human graph, including the model author's own, matched LTX 2's result — especially for non-human or unusual subjects.

Limits

  • 20 images is a local capability; the online graph caps at 5.
  • The NAG node is bypassed in the shipped graph because it errors online.
  • Control-video quality depends on keeping frame gaps small and the subject close to the target.
  • First-stage 0.67 scaling means the numbers you enter are not the output size.
  • The IC-LoRA Union result is explicitly less impressive than the multi-image graph.
  • For digital humans, LTX 2 remains the better tool.

Episode resources

Top comments (0)