Repository navigation
Releases
Atelier 0.18.0
What's new
- Agents can fix faces and edit parts of an image. New tools find the faces in an image (
find_faces), grow or shrink the mask (grow_canvas_mask) and frame a region at the zoom you ask for (set_canvas_frame), so an agent can mask a face, zoom in on it and regenerate it with more detail. MCP.md has the recipe. - Agents can select anything they can see. An agent looks at the image and selects what you name with the smart selection tool (
auto_mask), not only faces: "change her jacket to red leather", or "select the dog and give it a tiny raincoat". - Edits that leave the Create screen alone.
edit_imagedoes a whole masked, zoomed edit as one job in the Queue without touching your canvas, prompt or settings, so an agent can fix a batch of images while you keep working. - Agents can manage the moodboard.
get_moodboardandset_moodboardlist, add, remove, reorder and weight reference images. - Zoom to Fill. Beside Zoom to Fit in the View menu, the canvas menu and the zoom menu: enlarges the image until it covers the frame.
Atelier 0.17.0
What's new
- Paint on the canvas. Brush and Eraser now work on either the Mask or a new Color layer: switch with Mask | Color. Paint in color to guide what Atelier generates; with a mask, the paint only counts where the image is regenerated, and it blends in smoothly at soft mask edges.
- Paint stays editable. Your paint guides a run but never gets baked into the original picture. Painted images get a Draw entry in History that reopens on the unpainted picture with the paint still editable.
- Fill, Sketch and more. Fill a selection or the whole mask with color, start a New Sketch on a blank canvas, and pick recent colors from the color bar.
- Mask bar. Invert, Clear Mask, Copy and Paste Mask, and Mask Colored Areas (turns your paint into a mask) now sit in a bar over the canvas.
- Show or hide mask and paint. The eye button hides both; hidden layers don't affect a run. Each image remembers its setting, and a masked run's result starts with its mask hidden.
- Mask Opacity setting for how strongly the mask is drawn.
- Agents can paint. New
set_canvas_painttool, blank sketches, and paint and hidden-layer info in the canvas tools.
Fixes
- Fixed a rare hang when two saves to the same project overlapped.
Atelier 0.16.0
What's new
- Share your work. A new Share button sits beside Save and Export, and Share is in the right-click menu too, so you can send an image or video straight to Messages, Mail, AirDrop and more.
- LongCat-Video Avatar 1.6. The newest version of the talking-avatar video model. Atelier downloads, moves and deletes the extra file it needs for you.
- Smoother video playback. Clips play instantly from your project, and new buttons let you step through them one frame at a time.
- More control over MiniMax H3 Image. Use any sampler and shift setting, and make images as large as 24 megapixels.
- Lock the aspect ratio. A lock in the size settings keeps the shape of your image while you change its width, height or megapixels. Sizes no longer jump around while you're still typing.
- Delete a whole project from the project menu.
- Delete images and videos right from the Create screen, without switching to History.
- See which LoRAs made an image. History now lists the LoRAs an image used, with their weights.
- Start and end frames from the moodboard. H3 first-and-last-frame videos can take both frames from the moodboard.
- Import MOSS Image fine-tunes made in ComfyUI.
- Refresh in Models now picks up models and LoRAs you've added to the list by hand.
Fixes
- Switching back to the Create screen no longer pauses for a moment when you have many LoRAs.
- Deleting an image in History no longer leaves it showing on the Create canvas.
- LongCat videos have more consistent audio volume.
- If macOS's audio system stops responding, playing a clip no longer freezes Atelier; the clip plays without sound instead.
- Importing a damaged or incomplete model file can no longer crash Atelier.
Atelier 0.15.1
Install: download Atelier-0.15.1.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
Bug fix
- History crashed on macOS 26 when scroll bars are always shown (for example with a mouse connected, or Show scroll bars: Always in System Settings ▸ Appearance). The grid sized its thumbnails for the whole width of the window, scroll bar included, which sent the layout into a loop. It now leaves room for the scroll bar.
Atelier 0.15.0
Install: download Atelier-0.15.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
Input entries in History
History now records what went into your runs, as Draw Things does, so you can get back to the exact canvas, mask and moodboard you started from.
- Imported: an image you put on the canvas (pasted, dropped, opened, from Photos, or placed by an agent).
- Moodboard: a change to the moodboard, holding the whole board, weight-0 images too. A burst of changes is one entry.
- Cleared: the canvas was cleared.
- Mask: when you generate with a mask that isn't the one the latest entry holds.
- Entries show in the grid, the graph and the strip under the canvas with a badge. Click one to put its canvas, mask and moodboard back; the prompt and settings stay as they are. Show Inputs in the History toolbar hides them.
- Double-click an image to load everything it was made with, input and mask included. Use Settings now leaves the canvas as it is.
- Agents:
list_historylists entries withkind: "input", andget_imagecan show an entry's canvas image.
Guide: History ▸ Input entries.
Faster repeat runs
- MiniMax H3 and H3 Image keep the last prompt's encoding, so running the same prompt and references again skips the text encoder: an H3 Image rerun goes from about 25 s to 3 s, HyperFlow from about 62 s to 8 s.
- Weights Cache keeps other models loaded when canvas or moodboard images change; only FLUX.3 reloads, and only when references are switched on or off.
- When the same model file is in more than one models folder, Atelier loads the copy on an internal drive, which is faster.
Guide: Settings ▸ Performance.
Bug fixes
- MiniMax H3 Image gave a black image (or failed) when JIT Weights Loading was Always or Always (Partial). Those settings split the model file into a
-tensordatafile, and part of it was then read wrongly. Fixed for every JIT setting; a model that was already split works again without converting it again. - MiniMax H3 Image could keep the wrong conditioning when its guidance setting was changed.
- Fixed a possible freeze when clicking an image in the strip under the canvas after changing the moodboard; moodboard images are now saved without holding up the app.
Atelier 0.14.0
Install: download Atelier-0.14.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
ReCaVSR video upscaling
A third way to upscale a video, next to Fast (Apple Super Resolution) and Quality (FlashVSR+). ReCaVSR restores a clip with a large model built on Wan2.2, a few frames at a time and reusing what it made for the frames before, so the result stays steady. It's in Upscale… and Export Video… for videos (not Upscale Image).
- Full or 8-bit. Pick Full or 8-bit weights per job. 8-bit uses about 3–4 GB less memory and looks nearly the same.
- Getting it. The first time, click Download ReCaVSR (about 10.8 GB, into a ReCaVSR folder in your models folder). For 8-bit, then click Make 8-bit Copy, which writes a 5.7 GB copy in about a minute.
- Size and speed. Up to 1920×1088. On an M5 Max it takes about 0.6 s a frame at 1280×720 and 1.3 s at 1920×1080, and it uses a lot of memory: about 23 GB at 720p and 30 GB at 1080p with 8-bit (27 and 33 GB full).
- Mac only. It doesn't run on an iPad. On an iPad, choose a paired Mac that has ReCaVSR in Run on. A server running an older Atelier doesn't offer it either.
- ReCaVSR is by Xijun Wang and others at the University of Science and Technology of China, Apache-2.0; see Settings ▸ About.
Guides: Making videos ▸ Upscaling videos, Getting ReCaVSR, Running jobs on another device.
Zoomed refines keep their detail
Refine a region with the frame zoomed in past the image's own pixels (say 4× on a face) and the region now keeps the resolution it was made at, as Draw Things does, instead of being shrunk back into the image.
- The image keeps its size on the canvas and in History, and the canvas shows the extra detail when you zoom in on it. A masked refine keeps the detail only inside the mask; the rest of the image stays exactly as it was.
- A later refine of that image starts from the detail and keeps it.
- Show Image Full Size (⌘Y) shows the image at its finest detail. Save, Copy, drag, Photos and the output folder export at full detail too; the new Settings ▸ Images ▸ Export Refined Images At picks Full Detail (the default) or Image Size. Agents'
get_imagestays at the image's size unlessfull_detailis true. - Strength now says how many steps a run goes through, for example "≈1.4 of 8 steps at Shift 5", since Strength is a noise level and a low value at a high Shift can run only one pass. For HyperFlow and MiniMax H3 Image, which run their own fixed schedule, the line says so.
Guides: Creating images ▸ Refining part of an image, Settings ▸ Saved images, MCP.
Match Colors… sheet
0.13's Match Colors to Original stays one click, with the same defaults. Match Colors…, just below it in the canvas menu, opens a sheet that lets you choose how far it goes.
- It shows the original and a preview of the match. Side by Side puts them next to each other; Slider shows one picture with a divider you drag across it. The sheet can be resized on the Mac and opens at that size next time.
- Four sliders, Tint, Saturation, Brightness and Contrast, each running from Edit to Original. A line under each says how the original differs ("The original is darker."), and a slider with nothing to match is unavailable and says "Same as the original."
- Ignore edited areas compares the two pictures only where the edit left the picture alone, so a red coat the edit added isn't taken for the whole picture having turned red.
- The Queue row shows the settings used. The
match_colorsMCP tool takes the four sliders andunchanged_only;strengthandluminancestill work as shortcuts.
Guides: Creating images, Queue, MCP.
History
- Thumbnail size. A slider in the toolbar sets the size of the grid's thumbnails, with small and large picture icons beside it that step it down or up. Pinching the grid does the same, and on iPad it's the Thumbnail Size menu.
- Wider detail panel. Drag the edge of the detail panel to make it wider or narrower.
- Full-size viewer. Press Space or ⌘Y (Image ▸ Show Image Full Size) to open the selected image in the same viewer as Create, with the arrow keys moving through the selection. Videos still open in Quick Look.
Guide: History and projects.
Models, LoRAs and presets
- Sortable columns. Click a column header in Models or LoRAs to sort. The LoRAs screen remembers its search, family, selection and sort while you visit other screens, and Models remembers its sort.
- LoRA weights. A LoRA added in Create starts at the weight and slider range in its
custom_lora.jsonentry (as Draw Things' LoRA Config Editor script sets them) instead of always 1 with −1.5 to 2.5. - LoRA import scale factor. Each LoRA in Import LoRAs has a Scale field (default 1), baked into the converted file as in Draw Things' import.
- Presets for models you don't have. Under Other Models, a preset whose model isn't here can now be applied to the current model with Use with [model], as Draw Things allows.
- Preset import. Importing a
custom_configs.jsonagain no longer makes numbered copies. A preset with the same name and settings is skipped; one with other settings or another model asks Replace, Keep Both or Skip, with Do this for all remaining conflicts. Import from Draw Things uses the same sheet.
Guides: Models and LoRAs, Creating images ▸ Presets.
Bug fixes
- Deleting a video from History takes a fraction of a second instead of minutes in a project with many videos, including clips with refined frames. A 50-frame clip went from about 23 s to about 0.35 s.
- Clicking a video thumbnail no longer leaves the previous video playing on the canvas while the new one loads, and a slow earlier load can no longer replace a newer clip.
- Double-clicking a video in History shows the clip rather than the still image it was made from.
- Use This Frame as Input, Copy Frame and Add Frame to Moodboard are unavailable until the clip has loaded, instead of using its first frame.
- Deleting a step from a project that came from Draw Things could write the deleted step back over the step that moved into its place.
- Show Presets Folder and Show Layouts Folder now reveal the folder in Finder; they did nothing before.
- Replacing a preset with one for another model no longer leaves the other model getting the old values on Reset.
- Strength's line under H3 models no longer disappears while the model's schedule is read.
- On a zoomed refine, pixels you painted in the source image stay opaque under the refine's detail layer instead of showing through.
- A refine's merged edges at fractional zooms are no longer faded twice.
- Match Colors…: a flat or black-and-white edit no longer reads "more vivid" with a slider that does nothing, and something colorful the edit added no longer loses saturation when Ignore edited areas is on.
Atelier 0.13.0
Install: download Atelier-0.13.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
Match Colors to Original
Edit models (Fibo Edit, FLUX Kontext, Qwen Image Edit, the MOSS Edit LoRA), image to image and refines often come out a little warmer or cooler than the picture they started from. Control-click the edited image on the canvas and choose Match Colors to Original to put its color balance back, leaving its brightness alone.
- The original is the image the edit was made from, or for an edit made only from moodboard images, the first one. Matching a match again goes back to the same original.
- The result is a new image in History after the edited one, and the status bar says how much warmth or tint was taken out.
- It runs as a job in the Queue on this device, with progress, and saves into the image's own project even if you open another before it runs.
- Agents can do the same with the new
match_colorsMCP tool, which can also match lightness and set a strength.
Guides: Creating images, Queue, MCP.
MiniMax H3 Image
- A normal row in Models. MiniMax H3 Image now shows in the Models list like any model you haven't downloaded (under Available or All), with Install… in the row, its detail and its right-click menu. Once installed, the installed model takes its place.
- 4-bit S copy. The install can make a 4-bit S copy (about 12 GB) instead of 8-bit S (21 GB), for Macs with less memory, at some cost in detail. Writing the 4-bit S copy takes a few hours.
Guide: Models and LoRAs ▸ MiniMax H3 Image.
Bug fixes
- Large images, such as upscales, show on the canvas right away instead of after several seconds, and clicking a small image right after a large one no longer waits for the large one to load.
Atelier 0.12.0
Install: download Atelier-0.12.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
MiniMax H3 Image
MiniMax H3 Image makes images from a prompt and edits pictures from an instruction, in one model, in 6 steps.
- Guided install. In Models, the MiniMax H3 Image row above the list has Install…. It links to the two Hugging Face downloads (the 40 GB model and the 1.2 GB VAE), then converts them on your Mac: it imports the model, converts the VAE, makes the 21 GB 8-bit S copy, and by default moves the full-precision copy to the Trash. The text encoder (MiniMax H3's) then downloads if you don't have it. Importing the files yourself with Import… works too, VAE included.
- Sizes up to 16 megapixels. For this model the size popover goes up to 16 megapixels (4096×4096). With the 8-bit S copy on a 48 GB Mac, 2048×2048 takes about 40 seconds and 4096×4096 under 2 minutes.
- Editing. Use Edit with the picture on the canvas; moodboard images are sent as further pictures, called
<Picture 1>,<Picture 2>and so on in the prompt. - Tips. It works best with a long, detailed English description of the whole image. When editing a picture it made, use a different seed from the one that made it.
- License. The model's license is the MiniMax-H3-Image Community License, which doesn't cover use in the European Union, the United Kingdom, South Korea or the United States. Atelier doesn't distribute the model; read the license on Hugging Face before you download it.
Guide: Models and LoRAs ▸ MiniMax H3 Image.
Refining part of an image
- Zoom and mask. Zoom and move the image under the frame to work on part of it, and paint a mask with the Brush and Eraser in the status bar. With Merge into Image (on by default) the result goes back into the whole image at its own resolution, blended over Mask Blur; a frame reaching past the image grows it. Results are saved as Draw Things saves a canvas, as layers.
- Select. The Select tool makes the mask for you: Select Subject uses Apple's subject detection, and clicks and boxes use SAM 2.1, which downloads once (93.9 MB for Small). Add, Subtract or Replace puts the selection into the mask; a selection not yet added is added when you generate.
- Masks stay with their images. Each image remembers the last mask it had and gets it back when it returns to the canvas, also after a relaunch. Mask ▸ Replace Canvas Mask / Add to Canvas Mask in History, and Copy Mask / Paste Mask, reuse one image's mask on another. Mask strokes, selections and setting changes share one undo history.
- Agents can do the same through new MCP tools:
set_canvas_image,set_canvas_mask,get_canvas_mask,auto_maskandset_canvas_region, andgeneratewithuse_canvasruns exactly what the Generate button would.
Guides: Creating images ▸ Refining part of an image, MCP ▸ Masks and regions on the canvas.
Also new
- MOSS Image Edit LoRAs. LoRAs from the MOSS Image Edit trainer import like any other LoRA and turn MOSS Image into an edit model, with the canvas and moodboard as references. Single-image edits work well; the released LoRA was trained on single-image edits, so edits combining images usually return the first image unchanged. Guide: Models and LoRAs ▸ MOSS Image Edit LoRAs.
- Load Settings from Image works for images made in Atelier, on the canvas or picked in History, and for any video frame put on the canvas with Use This Frame as Input.
Atelier 0.11.0
Install: download Atelier-0.11.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
Faster MiniMax H3 videos
- Sol Attention. Draw Things' sparse attention for MiniMax H3, FastH3 and HyperFlow H3 has its own inspector section, under TeaCache. It's much faster on long or large videos, with a small quality cost. Start Step sets when it begins and Tau how much it approximates (higher is faster). It's off by default, as in Draw Things, and is saved with each video, in presets and in settings imported from Draw Things. Guide: Video ▸ Audio.
- Previews no longer slow generation. Step previews of a video decode on their own instead of holding up the next step. With High Resolution Preview on, that was about 4 seconds a step on a 1344×768 H3 video.
- Faster finish. Once a video is decoded, its MP4 is written from the frames still in memory and its thumbnails are made in parallel, so it's ready to play sooner.
Models across drives
With more than one models folder:
- Move To in the Models and LoRAs screens moves a model or LoRA to another folder, with the files it needs and its list entry. Across drives each file is copied and checked before the original is deleted. A Folder column shows where each one lives.
- Add To at the bottom of the Models screen, and in the import sheets, picks the folder that downloads, imported checkpoints and LoRAs, and 8-bit copies go to.
Guide: Models and LoRAs ▸ Moving between folders.
Fixed
- Generating with a LoRA right after a run without one no longer crashes with 8-bit models.
- Clicking a video in the recent strip selects that video, not a later one made from it.
- The distilled Ref2VA models (FastH3 Omni, Pika H3 Flex, HyperFlow H3 Flash Ref2VA) need at least one reference image, referred to in the prompt as
<Picture 1>; the docs now say so.
Atelier 0.10.0
Install: download Atelier-0.10.0.zip, open it, and drag Atelier to your Applications folder, replacing any older copy. If you have 0.3 or later, Atelier ▸ Check for Updates… installs it for you. Requires a Mac with Apple silicon running macOS 26 or later.
HyperFlow H3 Flash
HyperFlow H3 Flash is a compressed MiniMax H3 that makes a video with sound in 4 to 8 steps, in a first/last-frame (FL2VA) and a reference (Ref2VA) version. It isn't in the download list yet, so you set it up yourself:
- From its Hugging Face page, download a GGUF from
fl2va_gguforref2va_gguf(Q8_0 is about 15 GB), plus theconfig.jsonfrom the matchingfl2va_int8_convrotorref2va_int8_convrotfolder. - Convert it with hyperflow_gguf_to_safetensors.py (
pip install gguf numpyfirst):
python hyperflow_gguf_to_safetensors.py <file>.gguf hyperflow_h3_flash_fl2va.safetensors --config config.json
Keephyperflowin the name, and putref2vain it for the Ref2VA file. - In Models, import the
.safetensorsfile. Atelier recognizes it and sets 6 steps, 1344×768 and 124 frames at 24 fps. - Select it and click Make 8-bit S Copy, new in this release for any imported MiniMax H3 model. The copy is about half the size and needs half the memory.
Leave the step count at 4, 6 or 8, guidance at 1 and the negative prompt empty. The weights are under the MiniMax H3 Community License, which doesn't cover the EU, the UK, South Korea or the US without MiniMax's permission. Full guide: Video ▸ HyperFlow H3 Flash.
Fibo 2.0 Edit Turbo
Fibo 2.0 Edit Turbo from Bria edits an image from an instruction ("give the fox a scarf") and also makes images from a prompt alone, in 4 steps. Its license is CC BY-NC 4.0, non-commercial use only, and you accept it on Hugging Face before downloading.
It isn't in the download list, and setting it up takes developer tools: convert its Qwen3-VL text encoder with the engine's ModelConverter --fibo2-text-encoder, merge its two transformer files into one fibo_2_….safetensors, and import that in Models. Then put an image on the canvas, switch to Edit and describe the change. Full guide: Models and LoRAs ▸ Fibo 2.0 Edit Turbo.
Improved
- Generate with an empty prompt. Generate, ⌘↩, the Queue and the agent
generatetool no longer need text. - Videos wait for Play. A finished video stays paused on the canvas until you press Play, as in History.
- See where the time goes. Each run records how long loading, encoding, each step, decoding and saving took. The info panel and Queue ▸ Recent show it, and the agent tools return it.
- Decoding… while it decodes. The status switches from the last step to Decoding… once sampling ends, instead of sitting on "Step N of N" through a long video decode.
- Large images load with a spinner. A big upscale that takes a while to open now shows a loading indicator instead of the old image.
- Typed frame counts snap to the closest supported length when you press Return or leave the field, instead of on every keystroke.
- The bundled copy of the community model list is up to date, with Pika H3 Flex's new recommended settings.
Fixed
- MiniMax H3 videos no longer come out covered in a grid of small blocks when Tiled Decoding uses a large tile size. H3 models always decode in 256-pixel tiles now.