Node-based generative pipelines
Build ComfyUI-style graphs that produce outputs meeting constraints.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- ComfyUI · Invoke · Houdini
- Class GUI-need label
HY· Hybrid
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R06.11.01.01-T01 Build an SDXL + depth ControlNet graph and render at a fixed seed
Original instruction
In ComfyUI (local, models preinstalled), build a graph: load input/room.png, compute a depth map (Depth Anything preprocessor), SDXL base with ControlNet depth strength 0.8, prompt from brief.txt, seed 42, 30 steps, dpmpp_2m karras, CFG 6, 1024x1024. Save the image to output/room_sdxl and the workflow to workflows/room_depth.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clicktype· Text & forms- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Workflow JSON: required node classes linked (checkpoint→sampler, depth preprocessor→ControlNetApply), params exact. Output 1024x1024 PNG; re-running workflow reproduces it bit-exact; depth of output (same estimator) vs input depth Pearson ≥0.8.
Proposed partial credit
35% graph structure; 25% parameters; 20% output matches re-run; 20% depth adherence
R06.11.01.01-T02 Set up an img2img pass with a LoRA at fixed strength
Original instruction
In ComfyUI, create an img2img workflow on input/sketch_cat.png using SD1.5 checkpoint dreamshaper_8, LoRA watercolor_v2 at model/clip strength 0.7, denoise 0.45, seed 7, 25 steps. Save output to output/cat_wc_00001_.png and workflow to workflows/cat_img2img.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clicktype· Text & forms- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Workflow JSON: VAEEncode of input feeds KSampler latent, LoraLoader strengths 0.7/0.7, denoise 0.45, seed 7. Output re-run bit-exact; output dimensions equal input; SSIM to input in [0.45, 0.85].
Proposed partial credit
40% graph; 30% parameters; 30% output reproducible and in SSIM band
R06.11.01.01-T03 Inpaint only the sky region and leave the rest untouched
Original instruction
In ComfyUI, inpaint input/street.jpg using mask input/sky_mask.png to replace the overcast sky with 'clear dusk sky, orange gradient' (SDXL inpainting model, seed 11). Composite the result back so pixels outside the mask are identical to the original. Save output/street_dusk.png and workflows/sky_inpaint.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & click- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Outside mask (eroded 3 px): max abs diff ≤1/255 vs original. Inside mask: mean hue in orange range and luminance gradient top→bottom increasing; CLIP score for 'dusk sky' ≥0.24 on sky crop. Workflow uses mask-composite node.
Proposed partial credit
40% outside pixels identical; 35% sky replaced per prompt; 25% workflow saved and reproducible
R06.11.01.01-T04 Batch-generate product variants from a prompt CSV
Original instruction
In ComfyUI build a workflow that reads 12 rows from input/variants.csv (sku, prompt), generates one 1024x1024 SDXL image per row with seed = 1000 + row index, and saves each as output/<sku>.png. Save the workflow as workflows/variants.json and run it once.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & droptype· Text & formscode· Code & commands- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
12 PNGs named by sku, 1024x1024; PNG metadata seeds 1000–1011 matching rows; CLIP similarity of each image to its row prompt ≥0.25 and higher than to any other row prompt for ≥10/12. Workflow re-run reproduces 3 sampled outputs.
Proposed partial credit
25% 12 files named; 25% seeds; 35% prompt fidelity; 15% reproducible
R06.11.01.01-T05 Upscale to 4096 px without visible tile seams
Original instruction
In ComfyUI take output/hero_1024.png and build an upscale chain: 4x-UltraSharp model upscale, then a tiled diffusion refine at denoise 0.25 (tile 1024, overlap 128) to exactly 4096x4096. Save output/hero_4k.png and workflows/upscale_4k.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & click- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Output 4096x4096. Seam metric: gradient discontinuity along tile boundaries (every 896 px) ≤1.2x background gradient level. Downscaled to 1024, SSIM vs source ≥0.85. Workflow includes upscale model and tiled sampler with overlap 128.
Proposed partial credit
20% size exact; 40% no seams; 25% content fidelity; 15% workflow
R06.11.01.01-T06 Repair a broken shared workflow so it runs end to end
Original instruction
workflows/broken_portrait.json from a colleague fails in ComfyUI: a missing custom node (unavailable here), an SD1.5 VAE wired to an SDXL model, and a width not divisible by 8. Fix it using only installed nodes so it produces the intended 832x1216 portrait, keeping prompts and seed. Save as workflows/portrait_fixed.json and render.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clickdrag· Drag & dropvis· Visual judgment- Work product
WP05· Time-based media- Conditions
hidden· Hidden/stale state- Proposed difficulty
- 3 / 5
Proposed verifier
Fixed workflow executes headless via /prompt with 0 errors; all node classes are installed; VAE matches SDXL; output 832x1216; prompts and seed unchanged from original JSON.
Proposed partial credit
35% runs without error; 25% missing node replaced; 20% VAE fix; 20% size and prompts preserved
R06.11.01.01-T07 Transfer a painting style while preserving a face identity
Original instruction
In ComfyUI create a portrait of the person in input/face_ref.jpg (synthetic identity) in the style of input/style_ref.jpg using IPAdapter (style) plus a face-identity adapter, SDXL, 1024x1024, seed 5. Save 4 outputs (seeds 5–8) to output/stylized/ and the workflow.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropvis· Visual judgmentclick· Point & click- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
For each output: ArcFace cosine to face_ref ≥0.55; style distance (VGG Gram) to style_ref ≤ threshold calibrated on gold set and lower than to 5 distractor styles. 4 files, seeds in metadata.
Proposed partial credit
40% identity kept (10 points each); 40% style match; 20% files and workflow
R06.11.01.01-T08 Generate a seamless 16-frame animated loop
Original instruction
In ComfyUI with AnimateDiff installed, generate a 16-frame 512x512 loop of 'campfire flames, night forest' at 8 fps with seed 3, export output/campfire.webp (looping) and output/campfire.mp4, and save the workflow. The loop must not jump between last and first frame.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI · AnimateDiff
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clickav· Audio/video over time- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Both files 16 frames, 512x512, 8 fps. Loop seam: mean abs diff(frame15, frame0) ≤1.5x median adjacent-frame diff. Temporal flicker (mean luminance std over time) under threshold. CLIP 'campfire' ≥0.25 on middle frame.
Proposed partial credit
25% format; 35% seamless loop; 20% low flicker; 20% content
R06.11.01.01-T09 Package a reusable brand-background subgraph
Original instruction
In ComfyUI create a group node (subgraph) 'BrandBackground' that takes a product image and hex color, removes the background, places the product centred on a gradient of that color at 2000x2000 with 8% margin, and outputs the image. Expose only inputs image and color. Save as a template and use it in workflows/brand_demo.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clicktype· Text & forms- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Template file exists with exactly 2 exposed inputs. brand_demo.json executes on 3 test products: output 2000x2000, product bbox margins 8±1%, background corner colors within ΔE2000≤3 of the input hex gradient endpoints.
Proposed partial credit
30% subgraph with 2 inputs; 20% template saved; 50% outputs correct on 3 tests
R06.11.01.01-T10 Queue the variant workflow through the HTTP API (control pair of T04)
Original instruction
Control pair of T04. With ComfyUI running headless and workflows/variants_api.json (API format) provided, write queue.py that posts one prompt per CSV row to /prompt with seed 1000 + index, waits on /history and copies results to output_api/<sku>.png.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- ComfyUI
- Interfaces
api· API/SDK- GUI need · task setting
AP· API-sufficient- GUI need · ordinary access
AP· API-sufficient- Interaction primitives
code· Code & commands- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Same output checks as T04 on output_api/; server log shows 12 /prompt calls from queue.py and no browser session.
Proposed partial credit
40% 12 jobs queued; 40% outputs correct; 20% seeds
R06.11.01.01-T11 Outpaint a landscape to a wider canvas in Invoke
Original instruction
In Invoke (Unified Canvas), extend input/lake_1024.png from 1024x1024 to 1536x1024 by outpainting 256 px on each side with an SDXL model, keeping the original centre region unchanged and the horizon continuous. Save the merged canvas to output/lake_wide.png.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Invoke
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Output 1536x1024; centre 1024 region SSIM ≥0.98 vs input; horizon line detected in extensions within 6 px of original horizon height at the boundaries; boundary seam gradient ≤1.3x background. Invoke DB has the canvas session with bbox edits. Agent VM has no terminal, curl or Python, so Invoke's REST API is reachable only through the Invoke UI; the server access log is checked for non-UI origins.
Proposed partial credit
15% size; 35% centre preserved; 30% horizon continuity; 20% no seams
R06.11.01.01-T12 Build a regional-prompt workflow in the Invoke node editor
Original instruction
In Invoke's Workflow editor build a workflow that splits a 1344x768 canvas into left/right regions using masks/left.png and right.png, prompting 'snowy pine forest' on the left and 'tropical beach' on the right with a shared style prompt, seed 21. Save it to the workflow library as 'SplitScene' and generate one image.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Invoke
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clicktype· Text & forms- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Workflow library entry SplitScene: regional guidance nodes reference both masks with correct prompts. Output 1344x768: CLIP zero-shot on left crop prefers 'snow forest' over 'beach' and vice versa on right (margin ≥0.03). Agent VM has no terminal, curl or Python, so Invoke's REST API is reachable only through the Invoke UI; the server access log is checked for non-UI request origins.
Proposed partial credit
40% workflow structure; 15% saved to library; 45% regional content correct
R06.11.01.01-T13 Install a local model and file outputs into a named board
Original instruction
In Invoke, use the Model Manager to install models/juggernautXL.safetensors from local disk with VAE sdxl_vae, create a board 'Client-Moodboard', and generate 4 images from prompt.txt at 1024x1024 seeds 100–103 directly into that board.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Invoke
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
AP· API-sufficient- Interaction primitives
click· Point & clicktype· Text & forms- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Invoke SQLite: model record with path and default VAE; board 'Client-Moodboard' holds exactly 4 images whose metadata show the model, seeds 100–103, 1024x1024 and the prompt. Agent VM has no terminal, curl or Python, so the REST API and invokeai.db are reachable only through the Invoke UI; server access log checked for non-UI origins.
Proposed partial credit
35% model installed with VAE; 20% board; 45% four images with metadata
R06.11.01.01-T14 Wedge render 27 variations with Houdini TOPs
Original instruction
In Houdini, scene shots/vase.hip: build a TOP network that wedges 3 parameters (glaze roughness 0.1/0.4/0.8, light angle 30/60/90°, camera focal 35/50/85) and renders each combo with Karma at 960x540 to render/vase_r{rough}_l{angle}_f{focal}.png, plus a contact sheet render/contact.png.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Houdini
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & dropclick· Point & clicktype· Text & forms- Work product
WP05· Time-based media- Conditions
licensed· Licensed pro app- Proposed difficulty
- 4 / 5
Proposed verifier
27 PNGs with exact names, 960x540; per-file embedded/attr metadata matches name; renders differ as expected (roughness raises highlight spread metric monotonic). contact.png is 9x3 grid of thumbnails in wedge order. .hip contains TOP net with wedge node.
Proposed partial credit
30% 27 renders named; 25% parameter correctness; 20% contact sheet; 25% TOP net saved
R06.11.01.01-T15 Author a procedural building generator as a Houdini Digital Asset
Original instruction
In Houdini create HDA 'test::building::1.0' with parameters floors (1–30), width/depth (m), window spacing, roof type (flat/gable). Output packed geometry with groups windows, walls, roof and a primitive attribute floor_id. Save to otls/building.hda and an example .hip with a 12-floor gable building.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Houdini
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & droptype· Text & formsgeom· Precise geometry- Work product
WP04· 3D/spatial model- Conditions
licensed· Licensed pro app- Proposed difficulty
- 5 / 5
Proposed verifier
hython test cooks HDA at 5 parameter sets: building height = floors × 3.2 m ±1 cm; window count = floors × floor(perimeter/spacing) ±1 per floor; groups exist and partition prims; floor_id range 0..floors-1; gable roof has ridge apex. Parameter ranges enforced.
Proposed partial credit
20% HDA interface; 35% geometry scaling correct; 25% groups/attributes; 20% roof variants
R06.11.01.01-T16 Scatter rocks on gentle slopes with spacing and export USD
Original instruction
In Houdini scene terrain/hill.hip, scatter 500 instances of rock variants (rock_a/b/c.bgeo) onto the heightfield only where slope < 30° and outside road_mask, with minimum spacing 0.5 m, random yaw and scale 0.8–1.2, then export instanced USD to export/rocks.usd via Solaris.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Houdini
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & dropclick· Point & clickvis· Visual judgment- Work product
WP04· 3D/spatial model- Conditions
licensed· Licensed pro app- Proposed difficulty
- 4 / 5
Proposed verifier
Parse rocks.usd PointInstancer: 500 instances, 3 prototypes; at every instance terrain slope <30° and road_mask=0; min pairwise distance ≥0.5 m; scales in [0.8,1.2]; yaw distribution non-degenerate.
Proposed partial credit
15% count; 35% slope/mask constraints; 20% spacing; 30% scale/yaw and USD valid
R06.11.01.01-T17 Reconstruct a lost pipeline that reproduces a reference image exactly
Original instruction
The client approved output/approved.png but its metadata was stripped. notes.txt says: SDXL base, one of 3 listed checkpoints, seed in 9000–9010, sampler euler or dpmpp_2m, 28 or 30 steps, prompt in prompt.txt, 1024x1024. In ComfyUI find the configuration that reproduces approved.png and save workflows/approved_rebuilt.json.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
HY· Hybrid- Interaction primitives
drag· Drag & dropvis· Visual judgmentcode· Code & commands- Work product
WP05· Time-based media- Conditions
hidden· Hidden/stale state- Proposed difficulty
- 5 / 5
Proposed verifier
Execute approved_rebuilt.json: output equals approved.png with PSNR ≥45 dB (deterministic GPU settings). Workflow parameters equal the hidden ground truth (checkpoint, seed, sampler, steps).
Proposed partial credit
20% checkpoint; 30% seed; 20% sampler/steps; 30% exact reproduction
R06.11.01.01-T18 Generate a two-character scene matching given poses and roles
Original instruction
In ComfyUI produce a 1216x832 image of two characters using input/pose_two.png (OpenPose skeletons): left is 'knight in silver armor', right is 'red-robed mage', with regional conditioning so attributes do not bleed. Seed 77; save output/duel.png and the workflow.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropvis· Visual judgmentclick· Point & click- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
OpenPose re-detected on output: per-person keypoint PCK@0.1 ≥0.7 vs input skeletons. CLIP on left person crop: 'silver armor knight' > 'red robe mage' and reverse on right (margin ≥0.03). Workflow has ControlNet openpose and regional nodes.
Proposed partial credit
40% pose adherence; 40% attribute separation; 20% workflow
R06.11.01.01-T19 Cut out a product and composite it onto a template
Original instruction
In ComfyUI build a workflow that removes the background from input/sneaker.jpg, places the cut-out on templates/bg_studio.png at 2000x2000 with its bottom edge on y=1700 and horizontally centred, adds a soft contact shadow, and saves output/sneaker_listing.png plus a transparent output/sneaker_cutout.png.
- Platforms
mac· macOSwin· Windowslin· Linuxweb· Any OS (web)- Applications
- ComfyUI
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clickvis· Visual judgment- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
sneaker_cutout.png alpha IoU ≥0.95 vs ground-truth mask. Listing image 2000x2000; product bbox bottom at 1700±6 px, centre x 1000±6; shadow region darker than template by ≥8% luminance under the product.
Proposed partial credit
40% matte quality; 35% placement; 15% shadow; 10% both files saved
R06.11.01.01-T20 Generate a tileable PBR texture set with Copernicus
Original instruction
In Houdini 20.5+ Copernicus (COPs), build a network that generates a 2048x2048 tileable 'mossy cobblestone' texture set: albedo, roughness, normal (OpenGL) and height, all derived from shared noise so features align. Export to tex/cobble_{map}.png.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Houdini
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & dropvis· Visual judgment- Work product
WP05· Time-based media- Conditions
licensed· Licensed pro app- Proposed difficulty
- 3 / 5
Proposed verifier
Four 2048² PNGs. Tileability: wrap-around edge difference ≤1.1x interior adjacent-pixel difference on all maps. Normal map derived from height (finite-difference normals correlate ≥0.8). Albedo/height edge maps correlate ≥0.6.
Proposed partial credit
20% four maps exported; 40% tileable; 25% normal consistent with height; 15% maps aligned
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.