artfat-creatorandclaude-flow
ce95fe7099
Adds two capabilities and removes four outputs.
New:
* "Megapixels · aspect" size mode: pick a pixel budget + a named aspect
(or ratio_w:ratio_h), W x H hit that budget snapped to `multiple`.
* Optional image/mask resize to the computed BASE W x H, with
scale_method + crop. Passthrough None when nothing is connected.
BREAKING: outputs went 8 -> 6. mid_width/mid_height/final_width/final_height
are gone; image/mask took index 2-3. ComfyUI links by index, so existing
graphs using this node will need re-wiring. The upscale stages are still
computed and shown in `info`.
Fixes the live readout, which mirrors the Python compute in JS:
* it had no Megapixels branch, so SIZES[size] was undefined and the
readout silently drew nothing in the new mode
* snap() used Math.round, which breaks .5 ties upward, while Python's
round() breaks them to even: 1680/32 = 52.5 gave 1696 in the UI and
1664 from the node. Now ties-to-even in both.
Co-Authored-By: claude-flow <ruv@ruv.net>
ComfyUI-Artfat-Resolution
A tiny, elastic resolution picker for ComfyUI built for low-base + upscale workflows (Krea 2 / Selfism / H3), but works anywhere. Four ways to set the size, two upscale stages, an optional image/mask resize to the computed size, and a live readout so you know the final file size before you render.
No canvas, no dozen buttons — clean controls and a real-time readout.
Node: Artfat Resolution
Four modes (one size dropdown)
- Preset size — pick a ready
W×H(e.g.2:3 portrait · 512×768). Beginner-friendly. - Custom ratio — set
ratio_w:ratio_h+baseshort side (e.g.3:1banners). - Megapixels · aspect — set a target area in
megapixels+ anmp_aspect; the node solvesW×Hfor that pixel budget and snaps both sides tomultiple(e.g.0.98 MP@9:16→736×1312).mp_aspect = custom (ratio_w:ratio_h)uses theratio_w/ratio_hwidgets instead of the named list. - Manual W×H — type an exact
manual_w/manual_h. Not snapped — exact stays exact (e.g.1200×400).
mp_aspect list: 9:16, 2:3, 3:4, 4:5, 1:1, 5:4, 4:3, 3:2, 16:9, 21:9, custom.
Two upscale stages
upscale_1,upscale_2— decimal multipliers.1.0= off (true passthrough, no snapping).- The node computes the full chain base → mid (×up1) → final (×up2) — matches a two-pass image or video upscale pipeline, so you see the final size up front.
- Since v3 the mid/final sizes are not output sockets — they appear in
infoand in the on-canvas readout only.
Optional image / mask resize
- Connect
imageand/ormask(both optional inputs) and the node resizes them to the BASEwidth×height— not to mid/final. scale_method—lanczos(default),nearest-exact,bilinear,area,bicubic. Passed straight tocomfy.utils.common_upscale.crop—center= cover-crop keeping aspect;disabled= stretch to exactW×H.
Other inputs
multiple— snap sides to 8 / 16 / 32 / 64 (16/32 safe for Krea 2 / Qwen / H3). Ignored for the exact side in Manual mode.base— short side, used only in Custom ratio mode.batch_size— for the emptyLATENToutput.latent_type—SD3 / Flux / Krea 2 (16ch)orSD1.5 / SDXL (4ch). Only matters if you use thelatentoutput — match it to your model.
Outputs
| # | Name | Type | Notes |
|---|---|---|---|
| 0 | width |
INT |
base width |
| 1 | height |
INT |
base height |
| 2 | image |
IMAGE |
connected image resized to base W×H; None if nothing is connected |
| 3 | mask |
MASK |
connected mask resized to base W×H; None if nothing is connected |
| 4 | latent |
LATENT |
empty latent at base size (drop your EmptyLatentImage) |
| 5 | info |
STRING |
e.g. 512x768 -> 960x1472 -> 1920x2944 · 0.39 MP |
info lists only the stages that actually change the size, and always ends with the base
megapixel count. Wire it into a text node for a readout / filename.
The node also draws the base → mid → final chain live at the bottom.
Gotchas
- Breaking vs v2 — the outputs changed. v2 was
width, height, mid_width, mid_height, final_width, final_height, latent, info(8 sockets); v3 iswidth, height, image, mask, latent, info(6). ComfyUI wires outputs by index, so v2 workflows will come back mis-linked — re-wire the node after updating. image/maskoutputs areNonewhen the matching input is not connected. Only wire them downstream when you actually feed something in.- The image path uses
image[..., :3]— alpha is dropped, the output is always 3-channel RGB. scale_method = lanczoson a mask goes through ComfyUI's PIL lanczos path, which round-trips through 8-bit. For soft / precise masks preferbilinearorarea.- The
latentoutput is built asheight // 8, width // 8. In Manual mode with sides that are not multiples of 8 this silently floors — use a snapping mode if that latent feeds a sampler. - Megapixels mode snaps both sides with round-to-
multiple, so the realised area is near the requested MP, not exact. The value actually reached is printed ininfo. - The on-canvas live readout (
web/artfat_resolution.js) mirrors the Preset / Custom ratio / Manual math only — in Megapixels mode it draws nothing. The node itself still computes correctly; read the numbers from theinfooutput.
Install
Option 1 — ComfyUI-Manager (easiest)
Custom Nodes Manager → Install via Git URL →
https://github.com/artfat-creator/ComfyUI-Artfat-Resolution
Option 2 — git clone
cd ComfyUI/custom_nodes
git clone https://github.com/artfat-creator/ComfyUI-Artfat-Resolution.git
Then restart ComfyUI (a browser refresh is not enough — the node registers on server start).
No extra Python dependencies — uses torch and comfy.utils, both already provided by ComfyUI.
License
MIT © artfat-creator