Deep dive¶
The fine print behind the guide: limits you'll only run into in particular situations, and what to do when you do. None of it is needed to use the app.
Edit: what it can and cannot change¶
The basics are in the guide. This section covers what to do when an edit doesn't do what you asked.
"Another take on it" keeps the composition and redraws the detail. The layout, framing, pose and props stay; folds, hands, hems and the exact crop change. It runs at a fixed strength of 0.8. Lower strengths were tried and changed nothing you asked for (a dress colour, the time of day, the style), so there is no slider.
Change what's there can't add something new. Repainting keeps what is already in the rectangle, so asking for a dog on empty grass gives you grass. Use add something new instead, which sends the rectangle to the instruction editor.
Everything outside your rectangle is left exactly as it was, pixel for pixel. Inside it, you get whatever the editor drew, so a slight change in tone can appear inside the rectangle too.
The new object is drawn at the size of your rectangle. The editor only sees what is inside the rectangle, so if the rectangle is too small, the object will be cropped. Draw it around all the space the object needs.
Tall or wide rectangles work. The render matches your rectangle's shape up to about 4:1. Beyond that the editor gets a slightly wider frame, and part of what it draws may be cut off.
Big rectangles come back softer. The model repaints a window about 2.5 times the size of your rectangle at 1024px, so a rectangle over about 400px on its long side is repainted at lower detail than the rest of the picture. The dialog warns you. For a large area, edit the whole image instead, or use several smaller rectangles.
Word the change the way the model reads it. An anime model reads replace
the cats with birds as the tags "cats" and "birds" and may produce neither.
With a rectangle, write what should be there: black cat. Instructions like
"replace…" belong to the no-rectangle route, which uses a model trained to
follow them.
Editing an upscaled image¶
An edit without a rectangle renders at the edit model's working size, not at the size of your image. Editing a 4096px upscale with Qwen Edit gives you back roughly 1024px (or the same number of pixels in your image's shape, e.g. 1344×768 for 16:9). The dialog shows the output size before you start.
Ways around it:
- Edit first, then upscale. The simplest option and usually the best.
- Upscale or 2x fast the edited result to get the size back.
- Use a rectangle. Only the rectangle is repainted, so the rest of the image keeps its full resolution.
- Use another take on it and tick "repaint at N×N instead" to edit at the image's full size. Slower, and the dialog shows how much. An edit by instruction doesn't offer it: Qwen Edit at full size would take tens of minutes.
Not affected: Fix faces works on the faces alone, so the image keeps its size. It skips a face that is already large (common on an upscale), because repainting it would add nothing, and says so. More like this makes new images. Animate always renders at about 832×480, or 832×512 on MiniMax H3.
Upscaling with a different model¶
By default an image is upscaled by the model that made it. You can choose another in the picker next to the button, for a change of style or a faster model, but expect the fine detail to change: softer faces, blurrier foliage.
Why a pose change doesn't animate on FusionX¶
FusionX moves the pixels of the first frame. It has no idea of a body or skeleton, so no wording will make a standing person sit down. You'll get a smooth clip where they shift or take a step instead. Start from a still that already has the pose you want, or use MiniMax H3: asked to stand up from a sofa, its figure did, in one smooth motion with the camera tilting up to follow.
It also doesn't understand negatives, which is why "the camera stays still" makes the camera move: it acts on the words that are there.
Where an H3 clip's sound comes from¶
Draw Things' API hands back an H3 clip's frames but not its soundtrack. It does save every clip to its own history, sound and all, so the app reads the sound from there once the frames arrive. That only works when Draw Things runs on the same Mac as the app; on another machine, clips are silent.
A clip also comes out silent if its sound can't be found, for example when the clip is deleted from Draw Things' history while it renders, or a new Draw Things project is started meanwhile. The clip itself is unaffected. If macOS asks whether the app (or the Terminal running it) may access data from other apps, allow it: Draw Things' history is that data.
Tune: what each round renders and how it decides¶
Every round renders one scene, written the way the app would prompt the model, quality tags and negative included: a young woman in steel armour, loose dark hair and a brass lamp. Skin, loose hair and metal are where a wrong sampler or too high a cfg shows first, and one figure makes the count unambiguous. The model runs as it ships: speed-LoRA ticks off, and step counts you typed in the table ignored. The seeds are 1111 and 2222.
- Kind of model. Only for a model still on its kind's settings, where its architecture has more than one kind. For SDXL that means Illustrious, Animagine and plain SDXL. Each kind renders the scene the way the app would prompt that kind: its own quality tags, negative prompt, clip skip, sampler and cfg, at its draft settings, on two seeds. Picking the same kind on both seeds moves the model to it, and the later rounds are judged on that kind's prompt. Anything less keeps the kind it has.
- Sampler and cfg. The model's sampler and the other common one (Euler a and DPM++ 2M Karras), each at the model's cfg and about a third lower: 8 renders at the draft size and the final's step count. The crops are shown at native pixels, because shrinking a picture hides exactly what you are judging. Speckle at the higher cfg only is cfg burn. Speckle at both cfgs on one sampler means that sampler is wrong for the model. The round is skipped for a model at cfg 1, whose few-step recipe should be left as it is.
- Final size. Two square frames at the final size, plus one at the widest shape (16:9) at the model's shape budget. "More than one" on a square frame steps the final size down 256px and renders the round again. On the wide frame it steps the shape budget down 64px. After three tries, or at the draft size, it stops and keeps the last size, noting that it is worth a look by hand.
- Final steps. The final's step count against two thirds of it, on both seeds. The fewer steps are kept unless you pick the other on a seed.
A render that comes back blank, as one flat colour or as a grey refusal card can't be picked; it shows the reason instead. If nothing in a round came back, Render them again retries those cells.
Saving rewrites the model's entry in config/models.private.yaml. A
comment above the entry lists each round's outcome and says the settings were
picked by eye. A step count you typed in the table still wins over the saved
final steps. Discard throws the renders away and leaves the model as it
was. Renders are kept under tune/ in the data directory until you save or
discard.
The tag vocabulary in detail¶
The guide explains what the tag checks do. Here is what the suggestions are made of.
If a word isn't a tag but there is a tag that means it, the composer offers it. Type male and you're offered 1boy, male focus and multiple boys; female offers 1girl. You choose, because only you know which you meant.
A plain adjective gets the tags that use it. Danbooru tags what is long or open rather than the quality itself, so long offers long hair and long sleeves, and likewise open, bare, wide and holding.
Some words have no tag at all. These come back with nothing offered. Words
like beautiful, gorgeous and stunning are already covered: every
prompt to these models starts with quality words (masterpiece, best
quality), and adding more only weakens the words that describe the picture.
Quality words aren't Danbooru tags but the models were trained on them, so they're left alone.
Some tags have a preferred spelling. Danbooru treats mist as another name
for fog, and aerial view for from above. The composer suggests the main
spelling, and only swaps it for you when the whole phrase is the tag.
Unknown words are never deleted. "green raincoat" isn't a tag, but cutting it to "raincoat" would lose the colour, so it's left as written.
The tag list is in config/danbooru_tags.json. To refresh it from Danbooru,
run uv run python scripts/fetch_tags.py.
Prompt labels¶
Images from your own prompt are labelled orig; the LLM's prompts are V1, V2 and so on (or T1/P1 when tag-based and sentence-based models are selected together).
The numbers keep counting across the whole session, so a second Enrich continues at V4. That way a label always refers to one prompt, even after the composer shows newer ones.
When Draw Things goes away, or stops rendering¶
Jobs wait rather than fail. If Draw Things is closed or its API Server setting has switched off, queued jobs wait until it answers again.
Sometimes Draw Things keeps reporting success but returns no image, or starts drawing everything twice (two heads, a second horizon). The app notices: the header says "Draw Things not rendering" and the queue waits. Restart Draw Things and the queue carries on by itself.
If a job fails again straight after Draw Things recovers, the problem is with that job (for example, a size the model doesn't support), and the message says so.
If images suddenly look wrong on models that were fine earlier, restart Draw Things before blaming the model or your prompt.
When the LLM fails¶
You get a message with a Retry button and the reason: Ollama isn't running, the model gave an answer the app couldn't use, or it didn't answer in time.
The time limit is five minutes (llm.request_timeout in config/models.yaml).
A large model can take longer on its first request while it loads, so Retry
often works. If it keeps timing out, choose a smaller model. See also
Setup.