# LTX-2.5 First-Last-Frame Seed Hunt Workflow
A ComfyUI workflow for **FLF2V seed hunting** on LTX-2.5 distilled (22B), pinning start and end frames while previewing at 0.3 MP before upscaling to 1.2 MP.
## How it works
**STEP 1 – Seed Hunt (4 previews)**
- Load a **first frame** (start image) and **last frame** (end image)
- Queue four times with different noise seeds; each produces an 8-step video at 0.3 MP (736×416)
- Stage 1 anchors both endpoints at strength 0.7, so they guide but don't lock the content
- Preview time: ~40–50 seconds per seed
- Compare transitions without upscale overhead
**STEP 2 – Final Render (chosen seed)**
- Copy the winning seed into the red **CHOSEN SEED** node
- Un-mute **SAVE FINAL**, mute the green seed-hunt group
- Stage 1 re-runs at 0.3 MP, stage 2 upscales the latent ×2, re-anchors both frames at strength 1.0 (for sharpness), and runs 3 refinement steps → **1472×832 (1.22 MP)**
## Key features
- **Shared conditioning + guides** — both endpoints and their strengths are set once, reused by all four previews
- **KJNodes Set/Get wiring** — clean routing; rename a Set and every Get follows
- **No prompt enhancer** — prompt feed is direct
- **Frame-aware preprocessing** — both start and end images resize to the hunt resolution before encoding
- **Two-level guide strength** — stage 1 uses 0.7 (soft anchoring), stage 2 uses 1.0 (sharp re-inject on the upscaled latent)
## What you need
- LTX-2.5 distilled checkpoint ltx-2.5-22b-distilled-transformer-bf16.safetensors)
- Video VAE ltx-2.5-video-vae-bf16.safetensors)
- Audio VAE ltx-2.5-audio-vae-bf16.safetensors)
- Spatial upscaler ltx-2.3-spatial-upscaler-x2-1.1.safetensors)
- Text encoder CLIP gemma4-12b-with-proj-ltx-2.5-bf16.safetensors)
- ComfyUI node pack: **comfyui-kjnodes** (for Set/Get nodes)
## Tips
- **Start and end frames should share aspect ratio** — the first frame's resize defines the empty latent dimensions
- **Keep duration at 5s or less** — FLF2V works best as a constrained reframe; longer clips invite the model to invent content in the middle
- **Add No dialogue to your prompt if there are faces** — LTX-2.5's audio branch animates lips when mouths are in frame, which breaks the anchor frames
- **Try 6–8 seeds** if the transition feels off in all four previews (e.g., pulls back too early)
- If all previews show the same failure, tweak the prompt add "the transition happens immediately", add "the camera holds the close shot for the first second") and hunt again
Best for: smooth reframes between two fixed keyframes, where prompt and timing matter and the detail pass can use a stage-2 refine.