Replicate Video Request

Same job as this pack's FAL video nodes, different provider: this one calls Replicate's hosted API instead. The README frames it plainly - "generate videos using Replicate's model API, with Wan2.2 as the default video model" - and the default model string, wan-video/wan-2.2-i2v-fast, confirms it: this is Wan 2.2 image-to-video, running on Replicate's infrastructure rather than your own GPU. If you already have a Replicate account for other work, or you've found their pricing or queue behavior preferable to FAL's, this is the equivalent node aimed at that provider instead.

Inputs that matter: prompt is required and describes the motion/content. image is optional and explicitly documented in its own tooltip as being "for the image to video only" - meaning this node can also run in a pure text-to-video mode if you leave image empty, using aspect_ratio (16:9 default, or 9:16) instead of an input image to set framing. num_frames (default 81, ranging 81–121) and resolution (480p or 720p) set your basic output shape; frames_per_second (default 16, range 5–30) sets playback speed. go_fast (default true) is the speed/quality trade lever, matching the "fast" suffix on the default model string.

The two LoRA fields are the standout feature versus this pack's other video-request nodes: lora_weights_transformer and lora_weights_transformer_2 let you point at LoRA weights to apply to Wan's transformer during generation, each with its own lora_scale_transformer/lora_scale_transformer_2 (0–3, default 1). That's a real capability gap versus a plain hosted-model call - you're not stuck with the base model's style, you can layer in a trained LoRA the same way you would locally, without owning the GPU that trained or normally runs it. sample_shift (default 12) is a Wan-specific sampling parameter worth leaving at default unless you know what you're tuning it for. disable_safety_checker defaults to true here, same pattern as this pack's FAL nodes - worth knowing rather than assuming filtering is on by default.

Outputs: video, width, height, fps, url - consistent with the other video-request nodes in this pack, which makes them reasonably interchangeable in a graph if you want to A/B FAL versus Replicate output for the same prompt.

Install: search "ComfyUI-utils-nodes" in ComfyUI Manager, or git clone https://github.com/zhangp365/ComfyUI-utils-nodes into custom_nodes and restart. You need a Replicate API token from your Replicate account, passed via the api_key input.

Where people get stuck:

• LoRA weights here need to be weights Replicate's hosted Wan 2.2 endpoint actually accepts - typically a URL or identifier Replicate recognizes, not a local .safetensors path the way a LoraLoader node would take. Check Replicate's model page for the exact format it expects before assuming any LoRA file you have locally will work by just pasting a path into the field.
• num_frames bottoms out at 81, not 1 - you can't ask this node for a short 2-second clip; the minimum frame count is baked into the field's range.
• timeout defaults to 300 seconds, same as this pack's other API-video nodes - raise it (up to 3000) before assuming a longer generation has failed rather than just being slow.
• Every call is metered by Replicate. If you're testing LoRA scale values, iterate in small numbers of runs rather than sweeping a wide range blind - each generation is a real charge, not a free local render.