Osmium Workshop v2.0.0

Osmium Workshop 2.0.0 and Comfy Bridge 2.0.0.

Osmium Workshop

• Transfer tags: right-click an image (or use its 3-dot menu) and pick Transfer tags to copy some or all of its tags onto other images. Untick the tags you don't want, pick the target images from the grid, and it adds them (undoable). Tick Also swap image names with one target picked to trade the two images' file names as well.
• Bucketing copies tags onto the originals: Bucket Images and Trainflow's Start now overwrite each original's tags with the tags of its active image, even when every image is already bucketed.
• Quicktags take less room: in Single view, a category's checkboxes sit side by side when they fit the panel, each only as wide as its label, instead of one long list. Right-click one of your own quicktags to edit or delete it.
• Fixed: deleting an image for good now deletes its original in original_images/ too. Before, the leftover original made a later Rename all fail partway, and the next bucketing brought the deleted images back with their old tags while wiping the tags of some you kept. Rename all also moves a leftover original out of the way when it sits on a number it needs.
• Fixed: after disabling images in a bucketed dataset, Trainflow's Start (and Bucket Images) re-made the disabled images from original_images/ back into the dataset, without captions, so Start stopped with "N images have no .txt caption". Disabled images now stay disabled.
• Single view is now where Sequential lives (desktop): Quicktags on the left, the image and its tags in the middle, and the Sequential controls in the right panel. Every tick applies to the image right away (undoable); step through with Prev / Next. The button at the top of the right panel switches to the Gallery tools and back. Tag Overseer is only in the gallery views (its tab switches Single view back to Grid). "Sequential from first" is gone, and "from selected" and the 3-dot "Sequential from here" are now Single from here.
• Quicktags can be reordered: drag the ⋮⋮ handle on a quicktag (within or between categories) or on a category header. Built-in ones too.
• Sequential mode's Perspective: "Front" is now Straight-on (writes straight-on), plus a new Facing viewer (facing viewer) and Looking at viewer (looking at viewer). "From front" is retired as ambiguous; images that already have it keep it. The 🧭 Perspective indicator counts the new tags.
• Windows 7 theme (legendary, 750 Edibits): the Aero desktop. A glass title bar with its gloss line, Explorer's command bar whose tabs light up as blue glass, grey two-tone buttons that take the blue hover quickly and let it fade slowly, a default button that gently pulses in dialogs, green progress bars with a running glint, glass-framed dialogs, Harmony-blue behind the gallery, and the boot screen's four coloured lights in the opening flourish. It uses Segoe UI on Windows and Selawik (Microsoft's open look-alike, now bundled) elsewhere.
• Frutiger Aero theme (epic, 360 Edibits): glossy aqua jelly buttons, grass-green active states, and a sky-over-hillside gallery. In night mode the buttons' gloss turns over. Settings ▸ Frutiger Aero background ▸ Choose image… sets any picture as its gallery wallpaper.
• Both themes' gallery backgrounds are wallpapers, so Scroll the gallery background leaves them still.
• Trainflow tab (Windows): trains a LoRA for Anima on the dataset you have loaded, using an Anima-TrainFlow folder and your own Anima files. Get the trainer from this release (Anima-TrainFlow-Osmium-2.0.0.zip): unzip it and run its Install_Requirements.bat once, which downloads the Python packages (about 5 GB) at the versions Trainflow is tested with. Start Trainflow buckets the dataset and starts training in the background. It keeps going if you close Osmium (which warns you first), and the tab reconnects to its progress, log, previews and checkpoints when you come back. Trains on Janima too, not just base Anima. An Advanced section opens up the trainer's settings for those who want them (network alpha and dropout, training the text encoder, precision, LR schedule, timestep sampling, VRAM options and more); left alone, it trains exactly as before. Clear training cache deletes the latent and text-encoder caches the trainer keeps in the dataset, which it would otherwise reuse even after you change images or tags.
• LoKr and LoHa in Trainflow, next to LoRA. LoKr packs much richer changes into a quarter of the file at the same rank, which is popular for art styles. LoRA+ (a ratio in Advanced) makes one half of each layer learn faster, often reaching the same result in fewer steps. Also new: more optimizers (Prodigy + ScheduleFree, AdamW ScheduleFree, Lion8bit, Adafactor), per-block torch.compile, checkpoint offloading to save VRAM, split attention, training the modulation layers, tag shuffling and tag dropout, a validation split with its own validation loss, and extra image folders trained alongside the dataset.
• Newer trainer: the bundled Anima-TrainFlow trainer is updated to sd-scripts v0.12.0.
• Continue training: if a run stopped short of what you wanted (step 600 isn't quite there), pick one of its checkpoints and train on from it instead of starting over. Any checkpoint can be trained further from its weights, saved under a new name. Turn on Resume state and later runs can also be resumed exactly, optimizer and learning-rate schedule included, with the step count carrying on.
• Bucket Images keeps an original of every image in original_images/ (Initial State is Originals again, and initial_state/ folders are merged into it), and bucketing again re-makes the copies from the originals, so you can try different dimensions as often as you like.
• Bucket sizes are remembered per dataset and start at 512 to 1024 (step 64). Set them in the Bucket Images dock or right in the Trainflow tab; both show the same values. A Training size picker sets them for you: ≈512 (fast, less detail), ≈640 (the default) or ≈1024 (more detail, slower), with a line saying what a 2:3 and a square image will train at.
• WD14 review window, rebuilt: tagging a batch in Tag Overseer with "Apply automatically" off now shows one card per image. Each card keeps its picture on the left and its own tag area on the right, with the tags WD14 added tinted so they stand out from the ones already on the image. Drop any tag to the "Dropped" row and click it to put it back, or type in more. A Tag sorting button groups each card's tags by category, and you can apply one image, apply all of them, or skip all.
• Edit Log upgrades: filter by edit type, flip between newest and oldest first, and page through the whole log. A bulk edit opens a list of every image it touched, each with its own undo/redo and a thumbnail you can open, and right-clicking an image in the gallery shows every edit made to that image, undoable one by one.
• Attention setting for Osmium Comfy (SynthDat Overseer): auto, Flash, Sage, xformers or plain PyTorch attention. Auto picks Flash Attention when it's installed.
• Delete several tags at once: Tag Overseer has a new Delete tags list with every tag in the dataset and how many images have it. Tick as many as you want and Delete from all takes them off every image in one go, and one Undo brings them back. Unlike Void it doesn't set up a standing rule.
• Curate tab: go through a folder of new images before they go into a dataset. See them all as a grid, or one at a time with the arrow keys, and click to zoom in. Select the keepers and add them to the open dataset (captions come along) or start a new dataset with them. Artstyles sort an artist's different styles into their own folders so they never get mixed in one dataset: Create Artstyle, then Assign to it. A red junk folder takes the images you're throwing out.
• Isolate (Split) (Windows, in the image card's Crop): cut a character sheet or a page of panels into separate images. Pick how many vertical and horizontal lines you want (mix them for a grid), drag each line where it should cut, or Reset lines to space them evenly. Every piece is saved as a new image with the original's tags, the original stays as it was, and one Undo takes all the pieces back out.
• Trainflow: retraining a trigger word starts a new folder (<name>_2, _3, …, with the LoRA named the same) instead of mixing into the last run's folder, so there's nothing to delete first. Continue training lists checkpoints from all of them.
• Trainflow: Send a checkpoint to a folder of yours (ComfyUI's loras, say) under a name you choose, with a favorite folder that's remembered. Nothing already there is overwritten.
• Fixed: Undo/Redo and the Log getting out of step. Undoing an older edit from the Log no longer wipes the tag changes you made after it; only that edit is reversed. After Rename all, Undo and the Log's buttons follow the images to their new names instead of acting on whatever image now has the old name. A step that was already undone from the Log no longer blocks Undo for older steps. Clearing the log no longer breaks undoing crops, rotates and splits. Undoing a Merge or Void now actually sticks: the tags it merged or voided are paused in its standing rule (which used to re-apply it on the spot), and Redo turns them back on.
• Fixed: Rename all could delete images. If a rename stopped partway, the files were left with temporary __dts_rename_tmp_… names, and running Rename all again could delete some of them. Rename all now checks every file first, puts everything back if it stops partway, and never overwrites or deletes a file by renaming it onto a name that's already taken. Revert bucketing no longer deletes the bucketed copy when the original can't be read.

Comfy Bridge

• The Windows 7 and Frutiger Aero palettes and faces.

• ControlNet tab: a pose editor for the reference image (desktop and phone). Extract pose (DWPose, needs the comfyui_controlnet_aux node pack), New pose or Edit pose from the reference section open it, and every edit becomes the reference image. Pose a mannequin (limbs, hands with fingers, head, feet; what Anima's any-test LLLite follows, since it reads the reference as a picture) or an OpenPose skeleton. Drag points (attached parts follow), drag the torso to move a person, click a body part to resize it for perspective (or use the mouse wheel), and shape the figure with Body width, Chest and Length. Mirror, Reset, more people, and the original shown dimmed behind the pose.

• Open folder (Windows): an Open button next to the output folder, and next to the Upscale tab's chosen folder, shows it in File Explorer.

• RES4LYF samplers: if the RES4LYF node pack is installed in your ComfyUI, its samplers (res_2m, res_2s, res_3s and the rest) appear in the Sampler picker after the usual ones, marked res4lyf:, along with its beta57 and bong_tangent schedulers.

• Gallery tab (Windows and Android): your generations in a big grid; click one to open it (zoom, arrow through the rest, see its settings). A folded Compare section above it lines images up side by side: Shift-click a gallery image (long-press on the phone) to add it to a slot, as many as you like. Click a slot to show it on the left, right-click to show it on the right. Side by side: click an image to zoom it, Shift-click it and its prompt slides up beside it while the pair moves over to make room. Overlay: the two images stacked; drag the slider under them to fade the top one in and out and spot the differences (or turn the live preview off and press and hold to see through instead). Animated throughout, with an Animations switch to turn the motion off.

• The gallery panel remembers where you were: closing it keeps the folder and scroll position.

• Upscale with USDU (Windows and Android): Ultimate SD Upscale as the other choice next to the normal upscale. It upscales the image you generate (the pass 2 image when 2-Pass is on), and any image you already have gets its own Upscale tab: choose the image, see it large beside the settings (click it to zoom in), and press Upscale this image. The result can go to your output folder or to any folder you choose. It redraws the image tile by tile with its own checkpoint, so you get real added detail. Every setting is there with a tooltip, and the tiles' prompt can be the image's own prompt (with or without your Global field), WD14 tags, or text you type. Needs the ComfyUI_UltimateSDUpscale node pack.

• WD14 tags (Windows and Android): tag the images your last generation saved, with a .txt caption next to each one if you want, or tag any image you pick and use its tags as your prompt. On Windows the .txt can go right next to the picked image. Needs the ComfyUI-WD14-Tagger node pack.

• Metadata tab (Windows and Android): open any image to see how it was made and which LoRAs it used. Each LoRA can be looked up on CivitAI and CivArchive (by its file hash when you have it, otherwise by name), with preview images, and added to your LoRA stack in one click when it's in your folder.

• LoRA training metadata (Windows and Android, in the Metadata tab): see what a LoRA was trained with, laid out like Xypher7's LoRA Metadata Viewer: base model, network type and dim/alpha, optimizer and learning rates, epochs and steps, training time, noise settings, the dataset folders with their image and repeat counts, and the training tags per folder with a filter and a "copy top 20" button, plus every raw key. Pick one of your LoRAs; on the phone your PC reads the file from its loras folders. On Windows you can also open any .safetensors file from disk.

• Download to PC (Windows, with Osmium Comfy): a looked-up LoRA can be downloaded straight into your ComfyUI loras folder, with a choice of mirrors, a folder browser and a new file name. Files are checked against the looked-up hash, and API keys for CivitAI and Hugging Face are saved under Download keys and only ever sent to their own site.

• Import from any image: Import Generation now reads images made outside the Bridge too (Automatic1111/Forge, NovelAI, InvokeAI, SwarmUI, Fooocus and other ComfyUI workflows), from PNG, JPEG or WebP. It shows the image's details first, with Import Generation and Use prompts only. A single prompt is sorted into the prompt boxes by tag type, and switching between one box and separate boxes carries the prompt over. Images from other tools bring their prompts and basic settings, never their models.

• Phone access can require a shared token (Windows): the "Let the phone app use it" row has an optional token field next to the port. Leave it empty for the old no-password LAN/Tailscale behaviour; set one and the phone must enter the same token (a field above its Test connection button) before it can use this PC. The status line says plainly when no token is set.

• Main LoRA is now optional. Leave it empty and generation runs normally with no LoRA, even on a machine that has none installed. The Main LoRA weight only applies when a LoRA is chosen.

• Friendlier default settings: CFG 4, 25 steps, 1024×1024, sampler Euler A, scheduler Simple, seed 0; pass 2 starts at seed 1 with 20 steps.

• Seed behaviour dropdown next to each seed: Fixed (the default) reuses the number you typed, Randomize rolls a new one each generation, and Increment / Decrement step it by one. The mode applies per queued generation, so a queued batch walks the seed.

• In-app Help (Windows): the ❓ Help button left of the Wiki button opens plain-language guidance for every setting, one section at a time, in the same layout as Osmium's Help panel.

• A Clear all prompt fields button empties every prompt box in one click, in either prompt mode, along with the Character trigger. The Negative box is left alone.

• Attention setting for Osmium Comfy, the same as in Osmium.

• Windows: you can reorganize a generated image from the gallery. Open its details panel in the lightbox and under Save details it shows what the image was saved under and includes the full File Namer options for the move, seeded from the image itself: reclassify as OC, pick a different character, or respec the filename (add the model, sampler, seed and the settings it was actually made with). Moves never overwrite anything already in the destination folder.

• "Use these prompts" next to "Use these settings" in an image's details loads only its prompt fields and leaves everything else as it is.

Comfy Bridge for Android

• The File Namer works on the Android app too. Under Generation settings, File saving chooses where phone generations go: an automatic rating folder, a Character folder (Tag list with Danbooru search suggestions, all OCs in one folder, or a custom name), and which parts the filename carries — the same scheme the Windows app uses, with a live preview of the full path.
• The phone can browse the PC's output folder (over Tailscale/LAN): a Phone | PC switch at the top of the gallery, with the PC side navigated like the desktop. The PC's grid loads small thumbnails and only an opened image is fetched at full size, so browsing stays light on data, and each opened PC image shows its generation parameters with "Use these settings" and "Use these prompts". With Persist Comfy on, it works even while the Bridge itself is closed.
• Find PC: a button next to Test connection searches your Wi-Fi network for a PC sharing Osmium Comfy and connects to it (it searches the port in the box beside it, 8189 by default).
• Recent connections: every address that connects is remembered with its token, so a dropped link is one tap to reconnect.
• Flaky links recover on their own: lost or silent requests are retried, and thumbnails load several at a time, so one lost reply no longer holds up the rest.
• Reference images pick reliably: the preview shows the picked image itself, re-picking the same image works, and a slow reply after an upload waits up to a minute before giving up. Upload failures say how much of the image had been sent.

Install

• Osmium Workshop (Windows): download Osmium.Workshop-2.0.0-win.zip, unzip it anywhere and run Osmium Workshop.exe. No installer; the app keeps its data beside itself.
• Trainflow (optional, NVIDIA GPU): download Anima-TrainFlow-Osmium-2.0.0.zip, unzip it somewhere with about 8 GB free and run Install_Requirements.bat in it once (it downloads PyTorch and the trainer's packages, about 5 GB). Then, in Osmium's Trainflow tab, choose that folder and pick your Anima (or Janima) DiT, Qwen3 text encoder and VAE.
• Comfy Bridge (Windows): download Comfy.Bridge-2.0.0-win.zip, unzip it anywhere and run Comfy Bridge.exe.
• Comfy Bridge (Android): download ComfyBridge-2.0.0-android.apk, open it on the phone and allow the install when Android asks (it is installed outside the Play Store). It updates over an earlier install and keeps your saved hosts and settings.
• ComfyUI node pack (for SynthDat and Comfy Bridge): unzip ComfyUI-DataSetManagerNodes.zip into your ComfyUI's custom_nodes folder. Unchanged since v1.8.0.