| --- |
| license: apache-2.0 |
| tags: |
| - comfyui |
| - workflows |
| - image-editing |
| - qwen-image-edit |
| - background-removal |
| - upscaling |
| pretty_name: Production ComfyUI Workflows (API format) |
| --- |
| |
| # Production ComfyUI workflows, API format |
|
|
| 38 ComfyUI workflows in **API format** (the shape `/prompt` accepts, not the |
| editor's save format). These are not demos — they are the graphs behind the |
| image tools on [prathom.ai](https://prathom.ai), running against a single |
| RTX 4080. |
|
|
| Everything here was debugged against real images. The notes below describe bugs |
| that were live in production and are easy to reproduce if you build the same |
| graphs yourself. That is the part worth reading. |
|
|
| ## The four bugs, and what they look like |
|
|
| ### 1. Background removal produces an inverted alpha |
|
|
| `RemoveBackground` (BiRefNet) returns a mask where **the subject is 0 and the |
| background is 1**. `JoinImageWithAlpha` wants the opposite. Wire them directly |
| and you get a transparent PNG of the *background*, with the subject cut out — |
| the exact inverse of the request. It looks plausible in a thumbnail, which is |
| why it survives review. |
|
|
| `InvertMask` between them fixes it. See `transparent-png.json`, node 4. |
|
|
| This applies to `JoinImageWithAlpha` only. `ImageCompositeMasked` takes the mask |
| the other way round, so `product-background.json` and `white-background.json` |
| correctly have **no** `InvertMask` — adding one "for consistency" breaks them. |
| The polarity is a property of the consuming node, not of the mask. |
|
|
| ### 2. Qwen-Image-Edit silently forces square output |
|
|
| `EmptySD3LatentImage` with literal `width`/`height` sets the output size |
| regardless of the input. A 1920×1080 photo through a graph with a hardcoded |
| 1024×1024 latent comes back square and stretched. |
|
|
| Two nodes fix it: |
|
|
| ``` |
| ImageScaleToTotalPixels(megapixels=1.0, resolution_steps=16) |
| ↓ |
| GetImageSize → EmptySD3LatentImage.width / .height |
| ``` |
|
|
| `resolution_steps: 16` matters — latent dimensions must be a multiple of 16 or |
| the sampler pads and you lose a strip of the image. See any of the 13 edit |
| workflows, e.g. `anime-converter.json`, nodes 12 and 13. |
|
|
| **Deliberately not applied to the 10 fixed-format workflows.** |
| `linkedin-banner` (1584×396), `x-header` (1500×500), |
| `youtube-thumbnail-generator` (1280×720) and the rest must emit their |
| platform's exact dimensions. Deriving size from the input would break the one |
| thing those graphs exist to do. |
|
|
| ### 3. A hardcoded seed means the retry button does nothing |
|
|
| Every generative workflow started with a literal seed. Someone who disliked |
| their cartoon and pressed the button again received a byte-for-byte identical |
| image, with nothing on the page to explain why. |
|
|
| The 20 generative workflows here use `"seed": "{{seed}}"`, and the caller |
| substitutes a fresh random value per submission. |
|
|
| **Four keep a literal seed on purpose**: `remove-scratches`, `remove-dust`, |
| `remove-noise`, `ai-face-restore`. Those are restoration passes, where two runs |
| disagreeing about the same physical damage is a defect rather than a feature. |
| If you would rather re-roll a restoration, change them — the trade-off is real |
| and it goes both ways. |
|
|
| ### 4. Compositing onto a fixed canvas ignores the source dimensions |
|
|
| `ImageCompositeMasked` with `resize_source: true` and a fixed `EmptyImage` |
| canvas scales the subject to the canvas rather than placing it on it. Portraits |
| come back squashed. |
|
|
| `white-background.json` takes the canvas from `GetImageSize` on the input and |
| sets `resize_source: false`. `ai-passport-photo.json` is the fuller version: |
| image and mask are scaled separately — `lanczos` for the image, `bilinear` for |
| the mask, because a sharpening filter on a mask produces ringing at the cut |
| edge — then recombined through `ImageToMask`. |
|
|
| ## Placeholders |
|
|
| These are templates. Substitute before POSTing to `/prompt`: |
|
|
| | Placeholder | Meaning | |
| |---|---| |
| | `{{input}}` | filename returned by `/upload/image` | |
| | `{{seed}}` | fresh random int per submission | |
| | `{{preset}}` | rescale factor. RealESRGAN emits 4x, so 2x → `0.5`, 4x → `1` | |
| | `{{instruction}}` | the edit prompt, for the Qwen graphs | |
| | `{{passportWidth}}`, `{{passportHeight}}` | target size in pixels | |
|
|
| Model filenames are literal and will need changing to match your `models/` |
| directory. |
|
|
| ## Models referenced |
|
|
| ``` |
| qwen_image_edit_2509_fp8_e4m3fn.safetensors 23 workflows |
| qwen_2.5_vl_7b_fp8_scaled.safetensors 23 |
| qwen_image_vae.safetensors 23 |
| birefnet.safetensors 5 |
| RealESRGAN_x4plus.pth 5 |
| ddcolor_paper.pth 2 |
| RealVisXL_V5.0_fp16.safetensors 1 |
| bbox/face_yolov8m.pt 1 |
| ``` |
|
|
| Tested on ComfyUI 0.30.1 / torch 2.12.1+cu130, RTX 4080 16 GB. The Qwen fp8 |
| graphs fit in 16 GB at 1 MP; they do not at 2 MP. |
|
|
| ## One operational note |
|
|
| If you run several of these back to back, check free VRAM **after** unloading |
| rather than before. Reading it once and refusing on a stale number rejects a |
| graph for a shortage that a `/free` call would have cleared. It presents as an |
| out-of-memory error while the GPU sits idle, and it is easy to mistake for a |
| hardware limit. |
|
|
| ## Licence |
|
|
| Apache-2.0, for the workflows. prathom.ai itself is not open source. |
|
|
