wan / v2.7 / fal / reference-to-video

Every video-maker knows the pain: the first frame nails your character, and by frame forty the face is someone else entirely. Reference-to-video exists to fix exactly that - you feed the model a reference of who or what should appear, and it holds the appearance across the shot. This node runs that mode on Wan 2.7, hosted by Civitai's orchestration API, with the job billed in Buzz and the result coming back as a native VIDEO in your graph.

There's a real irony in the model choice. The community's actual character-consistency champion in 2026 is Bernini, ByteDance's Wan-2.2-based editor, which does near-closed-source-level reference-to-video locally. Wan 2.7 itself has had a lukewarm reception. So why reach for this? Because Bernini needs a serious GPU and a fiddly ComfyUI setup, and this node needs a prompt and a browser tab. When you want consistency right now without the rig, this is the pragmatic path - and the reference features here are genuinely the interesting part of the mode.

How it works

The plumbing is identical to the other Civitai recipe nodes: your inputs go up as a videoGen workflow to Civitai's orchestration API, the fal provider runs the Wan 2.7 job, and the pack long-polls until the MP4 comes back as a ComfyUI VIDEO. You're buying capability, not privacy - images and prompts leave your machine.

The inputs that matter

prompt and cfg_scale (default 4) are required; everything else is optional.

• reference_images - up to a few images (max 20 MB each) that define the character/object appearance. Wire in LoadImage or feed URLs.
• reference_video_urls_json - reference videos for both appearance and motion (max 100 MB each). The tooltip has the key detail: use @Video1, @Video2, @Video3 in your prompt to point at specific references. That's the syntax that ties the prompt to the subject, so don't skip it.
• multi_shots - toggle for multi-shot output, i.e. holding the reference across several different camera setups.
• aspect_ratio (default 16:9), duration (1–30s), frame_rate, resolution (720p/1080p), seed, steps, negative_prompt, loras, and enable_safety_checker.

The outputs are the usual trio: video (the clip), workflow_id (for debugging and cost lookup), and raw_json (the full workflow response - useful when something fails and you want to know why).

Install and auth

Pack-wide install, once: ComfyUI Manager → Custom Nodes Manager → search Civitai Comfy Nodes → Install → restart. Manual path:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # just requests

Credentials: Civitai Auth node → CIVITAI_API_TOKEN → stored key/OAuth → browser login, in that order.

Where people get burned

The @VideoN syntax is the thing people miss - a reference video that's never mentioned in the prompt is just ignored context. Also: 20 MB per reference image is generous but not infinite, and video jobs are the priciest thing in the Civitai catalog, so multi-shot runs add up fast. And remember the moderation reality: whatever Civitai's fleet filters, your reference images get filtered too - this is not the tool for content the platform won't process.