// Use Retro Diffusion for pixel-art image generation, img2img edits, spritesheets, and animation experiments such as platformer walk cycles, turnarounds, and action sheets from reference images.
| name | retro-diffusion |
| description | Use Retro Diffusion for pixel-art image generation, img2img edits, spritesheets, and animation experiments such as platformer walk cycles, turnarounds, and action sheets from reference images. |
| metadata | {"short-description":"Retro Diffusion image and animation workflows."} |
Use this skill when the user wants to generate pixel-art images or animation sheets through Retro Diffusion, especially when the task involves reference-image-driven character work like side-view platformer walks, turnarounds, or action cycles.
Retro Diffusion does not expose a provider-agnostic model marketplace like fal. The important control is the prompt_style, and each style implies a specific asset contract:
The right way to use Retro Diffusion is:
input_imageBefore generating, ask:
prompt_style impose a fixed frame size?Core principles:
prompt_style is the mode selector, not just a flavor tweak.animation__four_angle_walking is a 48x48 workflow, animation__8_dir_rotation is 80x80, and advanced animations should match the starting frame size.input_image should be RGB with no transparency, and the prompt should describe the reference rather than assume the API will infer everything.return_spritesheet: true so downstream analysis stays deterministic.input_imagereference_imagesPOST https://api.retrodiffusion.ai/v1/inferencescheck_cost: truerd-pro-platformerrd-pro-editrd-pro-spritesheetrd-fast-character-turnaroundrd-plus-character-turnaroundanimation-four-angle-walkinganimation-8-dir-rotationanimation-walking-and-idlerd-advanced-animation-walkingrd-advanced-animation-idlerd-advanced-animation-attackPOST https://api.retrodiffusion.ai/v1/inferencesX-RD-Token: YOUR_API_KEYOutputs can include:
base64_imagesoutput_urlsbalance_costremaining_balanceImportant operational rule:
Animations normally come back as transparent GIFs. Add return_spritesheet: true when you want a PNG sheet instead.
For input_image:
data:image/png;base64, prefixFor side-view platformer walks:
rd_advanced_animation__walking first when you already have a neutral starting framewidth and height equal to that starting frameframes_duration deliberately instead of taking the defaultreturn_spritesheet: true was setFor multi-direction walking presets:
animation__four_angle_walking and animation__walking_and_idle are 48x48 workflows64x64 anchorFor eight-direction turnaround experiments:
animation__8_dir_rotation first when you want a one-shot directional sheet80x80animation__8_dir_rotation returns server errors or weak directions, fall back immediately to a staged rd_pro__edit workflowreference_imagesPrompt like animation direction, not concept art copy:
Good Retro Diffusion prompt components for character animation:
For advanced animation prompts in particular:
300 characters when possibleImportant live-use nuance:
side-facingsame costume and silhouettebow-butt melee attackno background clutterscripts/retro_inference_run.py
scripts/retro_experiment_matrix.py
scripts/prepare_reference_image.py
Keep project-specific artifacts in the user's working project, not inside the skill directory.
Good default layout:
experiments/retro-diffusion/configs/outputs/, artifacts/, or asset-staging directoryThe important rule is:
When a Retro Diffusion run appears to stall, timeout, or return an ambiguous result:
Practical rule:
❌ Anti-pattern: comparing incompatible animation styles as if they were the same task
Why bad: a fixed 48x48 four-angle walker and a reference-driven advanced walking sheet are not equivalent outputs.
Better: compare them as different Retro Diffusion strategies, not as the same contract.
❌ Anti-pattern: feeding transparent RGBA sprites directly into input_image
Why bad: the docs say input_image should be RGB with no transparency.
Better: convert the input to RGB first and keep the subject on a clean flat background.
❌ Anti-pattern: asking for “walk animation” without saying whether you want a GIF or spritesheet
Why bad: you may get a preview format that is harder to analyze downstream.
Better: request return_spritesheet: true when the goal is extraction or frame comparison.
❌ Anti-pattern: using verbose prompts with advanced animation modes
Why bad: the backend may internally expand the action text and hit a hidden 500-character validation limit.
Better: keep advanced-animation prompts minimal and literal.
❌ Anti-pattern: over-simplifying a prompt after a good run Why bad: deleting "extra words" often deletes the exact identity and motion constraints that were keeping the model on-style. Better: shorten carefully, but preserve the clauses that lock facing, silhouette, action type, and background behavior.
❌ Anti-pattern: retrying immediately because the wrapper did not print a clean success message Why bad: Retro Diffusion may already have produced the output sheet, and an unnecessary retry wastes time, money, and confuses source-of-truth selection. Better: inspect the target output directory and run artifacts first, then decide whether a second run is actually needed.
❌ Anti-pattern: assuming the reference image alone will preserve style
Why bad: animation__any_animation can still drift into the wrong move family or add inconsistent effects if the prompt stops reinforcing orientation and action semantics.
Better: pair the starting frame and references with a compact but explicit prompt that preserves orientation, identity, and action read.
❌ Anti-pattern: assuming a larger isometric anchor will work better Why bad: larger references can be less stable than compact prepared anchors, especially in advanced animation modes. Better: downscale the approved anchor to a compact square first, then try advanced walking.
❌ Anti-pattern: trusting animation__8_dir_rotation as the canonical turnaround path
Why bad: it can fail server-side or produce weak directional separation even at the documented 80x80 size.
Better: keep it as a cheap first probe only, and rely on staged rd_pro__edit when you need a dependable turnaround workflow.
❌ Anti-pattern: ignoring the built-in frame-size contracts Why bad: some styles silently clamp or ignore your requested size. Better: choose the style because its size/output format fits the task.
❌ Anti-pattern: treating Retro Diffusion as a general-purpose video model Why bad: this API is about pixel-art image and animation sheet generation, not free-camera video. Better: use it for sprite-native outputs and compare those against video-derived workflows later.
IMPORTANT: Do not converge on one Retro Diffusion mode for every sprite task.
RD_PRO, RD_FAST, and advanced animation styles based on the asset contractframes_duration deliberately for short attack vs longer walk testsreferences/api-and-styles.mdreferences/animation-workflows.mdassets/model-presets.jsonassets/prompt-profiles/Retro Diffusion is strongest when you meet it on its own terms:
RD Pro Edit for dependable isometric turnaround workBuild 2D browser games with Phaser 4: WebGL-first rendering, scenes, filters, lighting, shaders, DynamicTexture and RenderTexture, tilemaps, SpriteGPULayer, TilemapGPULayer, and Phaser 3 to 4 migration work.
Build 2D browser games with Phaser 4: WebGL-first rendering, scenes, filters, lighting, shaders, DynamicTexture and RenderTexture, tilemaps, SpriteGPULayer, TilemapGPULayer, and Phaser 3 to 4 migration work. Trigger: phaser 4, phaser v4, migrate phaser 3 to 4, phaser webgl renderer, phaser filters, phaser render nodes, phaser SpriteGPULayer, phaser TilemapGPULayer.
Design and implement lightweight Three.js (r150+) ES-module scenes—hero sections, interactive product viewers, particle backdrops, GLTF showcases, or quick prototypes—whenever prompts mention 'three.js/threejs scene', '3D web background', 'orbit controls', or 'WebGL demo'.