DOGMA v44 Semantic Source View

What it is

A resize node, honestly. DOGMASemanticSourceViewV44 takes an image and a target_side, and gives you back that image scaled so its long side matches the target, with both dimensions rounded to multiples of 16.

What makes it worth its own node in this pack is what the info string says about it: "source-only, no generated pixels." This is the pristine input plane - the master you're restoring, parked at the exact resolution the rest of the v44 pipeline plans, crops and prompts against.

Why a fixed working resolution matters

Semantic detailer pipelines get messy when resolutions drift. Your planner's VLM looks at something, SAM sees something else, the crop node resizes again, and by the time a region reaches the model nobody agrees what "this window" refers to. Pinning a single canonical source view removes that class of bug: SAM thresholds mean the same thing run to run, crop coordinates are stable, and comparisons between iterations are actually comparisons.

The 16-multiple rounding is a latent-alignment habit borrowed straight from the tiling side of the pack - 16 and 64 divisible dimensions are the only way to be sure a VAE and a patch grid aren't silently shifting your content by half a tile.

Also worth saying plainly: that number is the source plane, not the generation plane. You are not resizing to what the model will sample; crops get their own target resolution downstream.

Inputs and outputs

• image - the source. Only the first three channels are used.
• target_side - INT, default 3584, range 1536–4096, step 64.

Two outputs:

• image - the resized view, bicubic, with antialias enabled when it's a downscale (upscaling with antialias on is a blur machine, so the node switches it off above 1.0 scale).
• info - either v44 semantic SOURCE 6144x4096 -> 3584x2400 or, if the dimensions already match, v44 semantic SOURCE kept 6144x4096.

If the computed size equals the current size, the image is passed through untouched - no needless resample round-trip. That's the behaviour you want at the top of a chain you'll run a hundred times.

Where it goes in the graph

First. Feed your master in, take image out, and use it as the source for everything else - masks, planning, reference crops. Print info on a Show Text node if you're comparing runs at different target sizes, because it's the only record of what resolution you were actually working at.

At the default 3584 the node comfortably handles both directions: it will bring a 6K plate down to something SAM and a local VLM can chew on without choking, and it will bring a 2K plate up so that crop math and region grouping behave consistently across a mixed batch.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/axior/ComfyUI-DOGMA-Nodes

ComfyUI Manager → DOGMA Nodes also works. Restart, then look under DOGMA/v44 - note the category on this one is different from the rest of the semantic detailer block, which lives under DOGMA/Semantic Detailer. The pack's categories follow its version generations rather than its function, so search by name, not by folder.

No dependencies, no downloads, no keys. requirements.txt in the repo contains nothing but a comment.

Gotchas

• This is a real resample, not a metadata change. At 3584 from a 6K plate you are discarding detail before the restore pass has run. If your source is soft and small, upscaling here doesn't help anything downstream - the crop nodes will handle their own sizing.
• Aspect ratio is preserved; the long side is what target_side controls, so a portrait image lands at 3584px tall, not wide.
• Like the rest of this half of the pack, there's no tutorial and no community thread to compare against: zero reddit threads name DOGMA Nodes, and the README documents the WAN VACE prep nodes and samplers instead. Small node, small stakes - but read the source.