The C2C ▸ Sampler latent preview setting controls the live denoising preview that appears inside a sampler node. Its Auto mode picks the right previewer per model on every run, so SD/SDXL/Flux and Wan/video samplers all preview correctly with no manual switching.
Setting: Settings → C2C ▸ Sampler latent preview.
ComfyUI's own Auto preview resolves to Latent2RGB, which has no color factors for Wan/video latents — so a Wan sampler shows a blank preview. The usual workaround is to force TAESD everywhere, but that's overkill for image models. The C2C Auto instead chooses per model, and it does so inside the previewer resolver so it stays correct on every sampler callback (it survives ComfyUI's per-prompt preview reset):
| Latent type | Detected by | Previewer chosen |
|---|---|---|
| Video — Wan 2.1/2.2, HunyuanVideo, LTXV, Mochi, Cosmos | temporal latents (latent_dimensions ≥ 3), the taew*/taehv* decoder name, or the format's class name |
TAESD — Wan/Kijai samplers route this to their own video previewer; falls back to a Wan-factor Latent2RGB if the decoder file is absent (never blank) |
| Image — SD 1.5, SDXL, Flux, SD3 | everything else | Core Auto — TAESD when the decoder is present (sharp), else fast Latent2RGB |
| Option | Behaviour |
|---|---|
| Auto — smart, per model (default, recommended) | Chooses per the table above |
| Force On — Wan/video-aware (TAESD) | Always TAESD |
| Force On — fast, SD/Flux only (Latent2RGB) | Always Latent2RGB (blank for Wan) |
| Off | No sampler preview |
For a crisp video preview instead of the rough Wan-factor fallback, drop the
Wan approx decoders into ComfyUI/models/vae_approx/:
taew2_1.safetensors— Wan 2.1taew2_2.safetensors— Wan 2.2
(available from Kijai's WanVideo_comfy on Hugging Face). With those present, Auto's TAESD path shows a proper decoded video preview.
This is backend-only and purely additive. It does not draw anything,
does not touch any node, and does not overlap ComfyUI's own UI — it only
ensures ComfyUI's native in-node previewer runs, and picks the best method
for the model. The get_previewer wrapper is fully guarded: any change in a
future ComfyUI release simply no-ops and leaves core untouched, and the
previewer can never return None (so previews never silently stop).
Opt out entirely with the environment variable C2C_NO_FORCE_PREVIEW=1.
| Symptom | Cause / fix |
|---|---|
| Wan preview blank | Setting is on Latent2RGB — switch to Auto or TAESD |
| Wan preview rough/low-res | Add taew2_1/taew2_2.safetensors to models/vae_approx/ |
| No preview at all | Setting is Off, or ComfyUI launched with --preview-method none and the guard is disabled by C2C_NO_FORCE_PREVIEW=1 |
Existing install still on taesd |
The default changed to Auto for new installs; older installs keep the saved value — set it to Auto once |
Apache-2.0. Backend-only, no core overlap.