DEV Community

AI-KSK
AI-KSK

Posted on

Multi-image first/last frame with LTX 2.3 three-stage sampling: 4 images, 241 frames, 3 real steps

#comfyui #ai #tutorial

Keyframe injection is the part of LTX 2.3 that turns a video model into a storyboard tool. This episode pushes four images into one 241-frame shot through the KJ node, then renders in three passes — a low-resolution pass that establishes motion, and two refinement passes that pull the keyframe identity back into place. The author's summary of the division of labour is the clearest line in the episode: the first pass is responsible for movement, the last two are responsible for looking good.

Evidence note: the graph ships as JSON and is parsed below — node count, checkpoint, LoRA, sigma chains, latents, upscaler. The 5-image online limit (20 locally), the 0.10-0.15 megapixel rule, the frame indices and weights, the >700-frame degradation and the "limbs are still bad" verdict are all the author's reported findings in 字幕.srt.

Where this fits

You have several frames of the same shot — a storyboard, a character sheet, a sequence of poses — and you want one continuous clip that hits each of them. This is the graph for that, and it is also the graph that shows most clearly where LTX 2.3 still falls short: limb articulation.

The graph

Item Value
Nodes 76
Checkpoint ltx-2.3-22b-dev.safetensors
LoRA ltx-2.3-22b-distilled-lora-384.safetensors @ 0.5
First pass euler_ancestral_cfg_pp, CFGGuider 1, sigmas 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0
Passes 2-3 euler_cfg_pp, CFGGuider 1, sigmas 0.85, 0.7250, 0.4219, 0.0 (twice)
Latents 320x288x*97* frames; audio 97 frames @ 25 fps
Upscaler ltx-2.3-spatial-upscaler-x2-1.0.safetensors
RunningHub https://www.runninghub.ai/post/2033913591946027010/?inviteCode=rh-v1111

Fetched once: HTTP 200, title matches.

Step by step

  1. Load the guide images. The online graph accepts up to 5; if you update the KJ nodes locally you can push that to 20.
  2. Unify the sizes. All images are resized to match the first image. The first image is scaled to a 1536 longest edge, and because the rest follow it, you can compute the first-pass width/height from that single ratio.
  3. Place each image by frame index and weight. In the demo: 241 frames, indices 0 / 60 / 120 / 180, weights 1 / 0.5 / 0.5 / 1. Weight 1 on the first and last frame is deliberate; the two middle frames sit at 0.5. The author notes that leaving out the first/last weights also works, because with several references and three render passes consistency holds anyway — his third pass output is nearly identical to the second.
  4. Constrain the size in one node only — the megapixel scale node. He uses 0.10 MP and says 0.15 MP is fine if you have the performance, because a 4x upscale follows.
  5. Set the sigma chain. His custom sigmas amount to 8 actual sampling steps in the first pass.
  6. Second pass uses the upscale model's sigma, about 3 actual steps, and its job is only to pull structure and keyframe appearance back — not to invent motion.
  7. Carry the latent between passes with three identical nodes. This is the structural trick: three copies of the same node exist so the upscaled latent from the previous stage can be handed forward. Without them the pipeline cannot accept the previous stage's latent.
  8. Crop at the end. The KJ node does not produce exactly the frame count you asked for, so a crop aligns the output batch length with the 241-frame target. Audio is unaffected by the crop.

Frame counts and where they break

  • 241 frames is the demo setting and it holds.
  • 440 frames with 12 images was described as working well.
  • Beyond about 700 frames he reports degradation and says he has not solved it yet.
  • Prompts and guidance come from the earlier Wan 2.2 three-image first/last-frame workflow, adapted.

Honest assessment

  • The motion is good enough for scene transitions and character turns.
  • Limb articulation is poor — in the fight scene he shows, the arms and legs do not hold up. He attributes this to training, not to the graph, and notes Wan 2.2 handled limb motion better at least in the sense of not breaking.
  • He expects someone willing to tune parameters or restructure the graph to beat the Wan 2.2 tier quickly, especially given LTX's speed advantage over the older three-pass digital-human loops.
  • An audio caveat: audio is untouched by the pipeline, so it stays intact.

Limits

  • The online graph caps at 5 images; 20 requires a local KJ node update.
  • Frames beyond ~700 degrade, with no fix offered.
  • Limb motion is the weak spot and is framed as a model-training problem.
  • Only one node should control size; changing size elsewhere breaks the pass chain.
  • The crop is mandatory if you need an exact frame count.
  • Third pass adds little: it is nearly identical to the second.

Episode resources

Top comments (0)