ComfyUI Chibi Sprite Pipelines

Three saved ComfyUI workflows that turn a character LoRA into pixelated chibi game sprites — not single images, but repeatable frame sets for idle poses, walking, and running, each across four POV directions: left, right, towards the viewer, and away from the viewer.

The three workflows

Workflow	Purpose
Video_Game_Chars_Pixel_Chibi_Poses	Single-frame idle / key poses
Video_Game_Chars_Pixel_Chibi_Walking V2	8-frame walk cycles per direction
Video_Game_Chars_Pixel_Chibi_Running V2	Run cycles with the same POV routing

Each workflow shares the same core stack and differs in pose reference frames and motion-specific prompts.

Pipeline architecture

Model stack

• SDXL checkpoint (waiNSFWIllustrious_v140)
• Character LoRA (swap per subject — e.g. a character identity LoRA at strength 1.0)
• Pixel sprite LoRA (GAME_ElinSpriteNoobLocon_byKonan) for chibi / 8-bit style
• Output latent size 1024×1512, decoded through the checkpoint VAE

ControlNet conditioning

• ControlNet Union ProMax (diffusion_pytorch_model_promax)
• lineart_anime preprocessor (SDXL, 512px) via Art Venture nodes
• Pose guides loaded as reference images — standard game-animation keyframes: Contact, Going Down, Average / break, Going Up, mirrored for left vs right foot forward

Frame selection

• Multiple LoadImage nodes feed an easy imageIndexSwitch (or Comfyroll CR Image Input Switch) so one workflow covers every frame in a cycle without duplicating the graph
• Walking and running workflows chain switches to route side-profile pose guides vs away-from-viewer silhouette guides separately

Sampling

• KSampler: euler_ancestral, 40 steps, CFG 5, randomize seed
• ControlNet strength tuned per workflow (~0.5–0.7; walking/running use a shorter apply window — start 0%, end 20%)

POV-specific prompt engineering

Each direction gets its own positive / negative prompt bank — the model will mix signals if you reuse one prompt for all angles.

POV	Positive emphasis	Negative blocks
Towards viewer	Front orthographic view, both eyes visible, symmetrical face, motion toward camera	Side view, profile, 3/4 angle, cropped bust
Left / right	Side profile, one eye visible, motion in walk/run direction	Front view, both eyes visible, perspective tilt
Away from viewer	Back view, back of head/hair/clothes, legs stepping away	Face, eyes, mouth, chest, looking at viewer, 3/4 view

Shared positive tokens across all POVs: pixel art, chibi, 8-bit, solo, full body, white background.

Away-from-viewer fix (documented in-graph): A workflow note explicitly says to remove face / eyes / mouth / eyebrows from the character description when generating back-facing frames — otherwise the model fights the pose guide and renders a face on a back-view silhouette.

The hard part: away-from-viewer consistency

Side and front POVs stabilized once pose guides and per-direction prompts were split. Away-from-viewer was the outlier: frame-to-frame drift in silhouette, hair volume, and limb placement.

What helped:

• Dedicated back-view prompt + negative sets (block all facial features)
• Stripping facial descriptors from the character LoRA prompt on away-facing runs
• Separate away-viewer line-art pose references routed through CR Image Input Switch chains
• In the walking workflow, Florence-2 (CogFlorence-2.2-Large) captions away-facing pose guides to verify the preprocessor sees a true back profile (e.g. "side profile, facing away from the viewer") before generation

Running and walking V2 workflows both encode this as parallel prompt branches and switch routing — not a single one-size-fits-all graph.

Output

Per character + animation type, the pipeline produces a multi-frame sprite set ready for assembly into game-engine sprite sheets — idle keyframes plus directional walk and run cycles, each with left, right, toward-camera, and away-from-camera variants.

Why it matters

Generative pipelines need the same discipline as software: isolate variables (one POV per prompt bank), version graphs (V2 walking/running split from the base poses workflow), and encode lessons in the graph itself (sticky notes for hair prompts, away-view face stripping, which switch index maps to which frame).