Phantas 🎞

The hard part of a multi-shot video isn't making the motion - it's making the stills agree. Generate five keyframes independently and the character is a different person in each one; put those into a video and the shots morph between them instead of flowing. Phantas is the renderer that stops that. It takes a board from Phantas Storyboard, renders every keyframe on one seed and one style, and hands back the Morpheus shot chain that runs between them.

Why the frames actually stay consistent

A shared seed is the floor, not the mechanism - same starting noise, but the trajectories diverge within a few steps. What holds a board together, in order: the style block stamped into every prompt, then an anchor. The reference input picks the anchoring mechanism:

• off - text and seed only. Weakest, and on a keyframe chain it usually shows.
• redux - CLIP-Vision plus a Flux Redux style model (style_model and clip_vision inputs), applied as an attention bias. Works with any Flux dev model, Krea included.
• edit - the anchor VAE-encoded as a reference latent, the way Kontext and Qwen-Image-Edit read an input image. Strongest, but it means running one of those edit models.

In redux mode, reference_strength (0.6) dials the bias: 1.0 is full Redux, which tends to redraw its reference and make every frame the same picture; lower keeps identity while leaving the prompt in charge of the shot.

anchor picks which picture to anchor on: first frame (global identity, but neighbors can drift), previous frame (tight seams, but drift accumulates down a long chain), or first + previous (both, for one more encode per frame). A wired reference_image - a character reference or location plate - is added on top of all of them and is the only anchor frame 1 can have. first_frame lets you use a picture you already have as keyframe 1 instead of generating it.

The inputs that matter

• board - from Phantas Storyboard. The prompts are already written; this node just renders them.
• sampler_settings - the same Sampler Settings bundle Ouroboros and Chimera take. Chain two for a draft-then-polish pass on every frame (later stage denoise < 1). Its seed_mode decides whether frames share a seed; fixed (the default) is what keeps them related.
• width / height - match the aspect ratio you'll render the video at, because these pictures are also what the video model's writer looks at.
• cache: disk - rendered frames cached under a causal key. This isn't only a speed feature: ComfyUI's Cancel discards the whole run, so without a cache a stopped board loses every finished frame. With it, cancel is the stop button - re-run and the finished frames come straight back.
• redo ("3", "2-4", "1,4-6") - re-roll specific keyframes with their seed shifted by redo_seed_offset (same seed + same prompt would just hand back the same picture). Frames after a re-rolled one re-render too, because they were anchored to what it used to be; the ones before stay cached.

Outputs and where it sits

shots is the Morpheus chain with prompts left empty - that's what tells Morpheus Storyboard "these are yours to write." Neighboring shots share the same tensor for the frame between them, because it's one picture playing two parts. images, captions and settings_data feed Image Compare directly, and each keyframe is pushed to a Kinburg Live Log node the moment it decodes (live_preview on by default).

Installing it

Same pack: ComfyUI Manager → "Kinburg-Nodes", or git clone https://github.com/Kinburg/Kinburg-Nodes into ComfyUI/custom_nodes and restart. The models are the real download: whatever image model you're keyframing with, plus - if you want redux - a Flux Redux style model and its sigclip CLIP-Vision. No extra pip packages; the LLM half of the pack (which this node doesn't use) handles its own install.