Wan 2.2 I2V Fast (Replicate)

Wan 2.2 is the last open Wan, and it's still the best quality you can run locally - but "locally" is doing a lot of work. The 14B I2V model wants a serious card and minutes per clip, and the speed-LoRA shortcuts that make it usable are a whole fiddly sub-hobby of their own. This node skips all of it. You feed it an image and a prompt, it calls wan-video/wan-2.2-i2v-fast on Replicate, and a finished MP4 lands in your output folder. Your GPU sits there doing nothing, and so does your patience.

The name is the important part: this isn't the stock Wan 2.2 I2V, it's the Pruna Lightning step-distilled "Fast" variant. Where the full model takes ~50 denoising steps, the distilled one does its thing in about 8 - which is why Replicate can sell it by the call instead of the hour. Quality takes a small hit versus the unhurried local run, but for iterating on an idea or for animating a single reference image it's a completely reasonable trade.

How it works

Under the hood this node is a thin API client, not a sampler. It converts your IMAGE tensor to a PNG, uploads it (small images go as an inline data URI, anything over ~200KB via Replicate's Files API), fires replicate.run() with your prompt and settings, then downloads the resulting video and writes it to ComfyUI/output/video/ as replicate_wan-video_wan-2.2-i2v-fast_<timestamp>.mp4. That's the whole mechanism. No model files, no VRAM, no weights - just a pay-per-call bill.

It's one of 20 curated models in the Replicate Select pack, which is built Runware-style: a nodes.json catalog drives a code generator, and every node follows the same pattern. Auth is a REPLICATE_API_TOKEN environment variable; there's no key field in the node, which is honestly a good sign for a wrapper pack in a corner of the ecosystem where credential-sniffing malware already happened once.

The inputs that matter

Most of the schema is optional and you'll rarely touch it. The ones a beginner actually sets:

• image (required) - the first frame. Whatever you feed in gets animated forward. Feed a good image; Wan inherits its flaws.
• prompt (required) - what happens after the first frame. The model is better with motion verbs than with static descriptions.
• last_image - optional conditioning for the final frame, which makes transitions smoother if you're stitching clips.
• num_frames - 81 default, up to 121. The tooltip is blunt: 81 gives the best results.
• resolution - 480p or 720p. At 16:9 that's 832x480 or 1280x720. Start at 480p; that's where the price is kind to you.
• interpolate_output - set it and the node frame-interpolates to 30fps with ffmpeg on the way out.

Two things worth knowing because they're easy to trip over. First, pricing is pegged to duration at 16fps - bump frames_per_second and you're not really buying smoothness, you're buying a longer clip at the same frame count. Second, the lora_weights_transformer fields take raw .safetensors URLs from the internet. That's Wan 2.2's two-expert MoE showing up in the API: the high-noise transformer and the low-noise transformer_2 can be LoRA'd independently, matching how the local community applies speed LoRAs to only the low-noise pass.

Installing it

Install once for the whole pack. ComfyUI Manager → search "Replicate Select" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Restart ComfyUI. The only real dependency is replicate>=1.0.7, which Manager installs for you. There are no model downloads - that's the entire point. Then set your token:

export REPLICATE_API_TOKEN=r8_...

and restart so ComfyUI picks it up. If the node complains about a missing token, that's this exact step.

Where people get burned

The output is a file path (STRING), not video frames. Wire it into a VHS Load Video or just open output/video/ - you can't chain more image nodes straight off it. Also: your reference image and prompt leave your machine and go to Replicate, subject to their logging and their safety checker (you can disable it with disable_safety_checker, but the model's own limits remain). And remember the local-first reflex you see everywhere in this community cuts both ways: this node exists to not run locally, so check the per-call price before you iterate 40 times on one clip.