# LTX-2.5 Two-Stage Seed Hunt Workflow

A ComfyUI workflow for fast seed hunting on LTX-2.5 distilled (22B), using two-stage generation to preview at 0.3 MP (736×416) before upscaling to 1.2 MP (1472×832).

## How it works

STEP 1 – Seed Hunt (4 previews)

- Queue once with four different noise seeds; get four 8-step outputs at 0.3 MP in seconds

- Distilled model + low resolution = ~40-50 seconds per preview

- Compare composition, motion, and timing without the upscale cost

- Change seeds and queue again until you find one you like

STEP 2 – Final Render (chosen seed)

- Copy the winning seed number into the red CHOSEN SEED node

- Un-mute SAVE FINAL (Ctrl+M), mute the green seed-hunt group

- Stage 1 re-runs that seed at 0.3 MP, stage 2 upscales the latent ×2 and re-samples 3 more steps → 1472×832 (1.22 MP)

- Composition and motion carry over faithfully because stage 2 doesn't resample from scratch

## Key features

- KJNodes Set/Get wiring keeps routing clean — rename a Set_* and every Get_* follows

- Shared loaders — models load once, not four times

- No prompt enhancer — prompt goes straight to CLIP encode for clarity and speed

- Resolution flexibility — adjust width/height on the two primitives; stage 2 doubles both automatically

- Supports image-to-video via the bypass_i2v boolean (falses out the input frame for pure T2V)

## 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

- Preview layout is trustworthy — stage 2 upscales the same latent, so seed #1001 produces the exact same composition at both resolutions

- Duration and FPS matter — they cascade through the empty latent and frame count automatically

- Width/height must be divisible by 32

- Try 6–8 seeds before settling on one if the results feel close

Best for: rapid iteration on T2V/I2V prompts where composition matters more than final detail.