DEV Community

AI-KSK
AI-KSK

Posted on

Multi-image first/last-frame transitions in LTX 2.3 — the transition LoRA does the heavy lifting

#comfyui #ai #tutorial

LTX 2.3 can already chain several images into one shot through keyframe injection, but the joins between images tend to look stiff: the model snaps from image A to image B instead of moving between them. This episode is about one LoRA — a transition-enhancement LoRA — that makes those joins behave like camera movement instead of cuts, applied to a 4-image first/last-frame graph heading into a three-pass render.

Evidence note: node count, model and LoRA filenames with their weights, sigma chains, latents and CFG are parsed from the workflow JSON in 工作流及封面(视频215). The prompt-writing rules, the "10-15" first-pass size guidance and the LoRA-weight advice are the author's reported practice in 字幕.srt.

Where this fits

You have 2-4 stills that should become one continuous shot — a transformation, a costume change, a car turning — and you need the in-between motion to be smooth rather than interpolated. This graph is the current answer in the LTX 2.3 line, and it applies to more than first/last-frame work: the LoRA is described as useful in any LTX 2.3 flow where a scene change needs a better transition.

The graph

Item Value
Nodes 81
Checkpoint ltx-2.3-22b-dev.safetensors
LoRA ltx-2.3-22b-distilled-lora-384.safetensors @ 0.5
LoRA ltx2.3-transition.safetensors @ 1.0 — the transition enhancer
LoRA LTX2.3_Reasoning_V1.safetensors @ 1.0
First-pass sampler euler_ancestral_cfg_pp, CFGGuider 1
First-pass sigmas 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0
Later-pass sampler 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-x2-1.1.safetensors
RunningHub https://www.runninghub.ai/post/2039960294918721537/?inviteCode=rh-v1111

The RunningHub page was fetched once: HTTP 200, and the title matches the graph.

Step by step

  1. Set the basic controls: positive prompt, negative prompt, frame count, fps.
  2. Load four images. Four is what the online graph ships with; the author says to adjust the count locally as you like.
  3. Fill the two columns. The left column is when each image appears (frame index); the right column is how strongly it counts (weight). Getting these two right is most of the work.
  4. Run. No other mandatory tuning.
  5. Understand the render order. This graph has not left the image-to-video three-pass structure. The first pass renders small, then two further passes each double the spatial size — a total of 4x — which is exactly why the first-pass size has to stay small.

The prompt rules that actually matter

The author walks through eight rules, and says you can hand them to an LLM to write the prompt for you:

  1. Write the camera language first, then the subject and the scene.
  2. Emphasise the process of change — this is the core. How the subject changes, how the frame changes; be concrete.
  3. Cut the empty filler. No atmospheric word-salad.
  4. Add visual detail: light, material, colour.
  5. End with the trigger word the LoRA was trained on, placed late in the prompt.
  6. Keep the chronology explicit.
  7. Use the remainder as supplementation.
  8. The rules are not specific to this graph — they apply to text-to-video, image-to-video, multi-image first/last frame — any LTX 2.3 flow where you want better transitions.

What changed versus the previous graph

  • An LTX 2 NAG node was added so the negative prompt has effect.
  • The multi-image first/last-frame node is the functional core.
  • The official LTX 2x spatial upscale model is used, and RunningHub carries the 1.1 version.

LoRA weights: how far you can push

The VBVR reasoning LoRA (the subject of an earlier episode) becomes effective at 0.5. You can raise it to 1, and the ceiling is about 1.5. The author's recommendation: do not go above roughly 1.1; 1.0 is the balanced value. The same reasoning applies to the transition LoRA's placement: it is a single Load LoRA node, so inserting it into another graph is a one-node edit.

Results the author shows

  • A four-image "Rin" animation where the input images differ a lot — the output is described as smooth, without obvious stalls.
  • A seed growing into a sapling — also smooth.
  • A car transformation clip, used as the demonstration of what the model plus LoRA can do.

Limits

  • Results are still "somewhat luck-based": the LoRA raises the probability of a good pull, it does not guarantee one.
  • The first pass has to stay small (the author cites a 10-15 range in megapixel terms) precisely because the pipeline multiplies by four.
  • Four images is the graph's default; more inputs means more nodes to wire, and the author frames it as a local modification rather than a supported setting.
  • The JSON is an 81-node graph; the three-pass structure is inherited and not simplified.

Episode resources

Top comments (0)