ComfyUI · Graphic novel · 2026

AELITA: PuLID vs Flux Kontext for graphic novel storyboarding

Building an AI storyboard pipeline for a graphic novel adaptation of Alexei Tolstoy's Soviet sci-fi novel — and what the two main identity-preservation approaches actually do.

Tarkovsky — figures in a field, misty landscape

AELITA is a graphic novel adaptation of Alexei Tolstoy's 1923 Soviet science fiction novel — built entirely in ComfyUI on RunPod, using Flux-based models. The cast includes two human protagonists (Los and Gusev) and blue-skinned Martian characters with strict appearance rules across all 8 panels.

The central architectural question: which identity-preservation mechanism produces consistent characters across panels — PuLID (face biometric lock) or Flux Kontext (reference image conditioning)? The answer turned out to be neither simple nor symmetric.

PuLID anchors identity at the biometric level — facial geometry, skin tone, proportions. It is the right tool for photorealistic characters whose face must be locked across shots. But it fights illustrated styles: feeding it a drawn reference produces a photorealistic output, not an illustrated one. The style of the source image is discarded.

Flux Kontext conditions generation on the full visual content of a reference image — pose, clothing, color palette, spatial relationships. It does content/identity transfer, not style transfer. Like PuLID, it outputs in Flux's native photorealistic register regardless of whether the reference is illustrated or photographic. The implication: for illustrated graphic novel output, neither mechanism alone is sufficient.

The architecture that worked for AELITA: Flux Kontext fed with panel crops (not masked character extractions). No PuLID. No ControlNet depth. Prompts generated by Claude Sonnet API applying cinematographic rules — camera angle, LEFT:/RIGHT: spatial anchoring, scale anchoring, environment density minimums for wide shots. One 69-node graph for 8 outputs, JSON built programmatically by a dedicated generator script.

The open empirical question: a hybrid pipeline — PuLID vault generation on RunPod feeding Kontext Max per panel — is pending a 10-panel test. The hypothesis: PuLID generates photorealistic character reference frames; Kontext uses those frames as conditioning anchors while accepting a stronger style LoRA. If the hypothesis holds, illustrated output with consistent identity becomes achievable.

COMPARISON

DIMENSION	PuLID	Flux Kontext
Identity mechanism	Biometric face lock	Full-frame visual conditioning
Style preservation	None — overwrites with photorealistic output	None — same photorealistic register
Character masking	Required — one node per face	Not needed — panel crops work
Best for	Photorealistic film storyboard	Illustrated graphic novel (with style LoRA)

OPEN SOURCE · TOOL

ComfyUI — node-based Stable Diffusion UI

github.com/comfyanonymous/ComfyUI