v0.1.13

Fixed

• S3 output downloads — RunPod's S3-compatible API does not support presigned URLs, so large outputs (>10MB, offloaded to the network volume) silently failed to download and jobs appeared successful with no files. ForgeHub now signs the GET itself (SigV4) using the configured volume credentials.
• Silent output loss — output download failures are logged and reported as job errors instead of being swallowed; a job that produced zero downloadable files now surfaces an error instead of an empty success.
• WAF-blocked downloads — RunPod's edge rejects the default Python-urllib user agent (error 1010); requests now send an explicit user agent.

Full Changelog: v0.1.12...v0.1.13

v0.1.12

Added

• Update checker — the app checks GitHub Releases at startup and shows a banner with a direct platform-matched download when a newer version exists; manual "Check for updates" row in Settings.
• Guided single-box prompt (sectioned_prompt rework) — one text area pre-filled with --NAME-- section markers (SUBJECTS / LOCATION / SUMMARY / SCENE / MUSIC / SOUND) plus a legend strip with per-section hints; click a legend chip to insert a missing marker.
• prepend flag on prompt-target params — the style directive is injected at the top of the prompt, before the sectioned template.

Changed

• Prompt template uses --LOCATION-- (alias of ENV) and includes --SUMMARY--; the previous per-section boxes are replaced by the single guided text area.

Full Changelog: v0.1.11...v0.1.12

v0.1.11

Added

• sectioned_prompt recipe param — the MiniMax H3 R2VA prompt is now a fixed-label template: SCENE / ENV / PROP / VFX / MUSIC / SOUND boxes (labels can't be deleted), each with wildcard autocomplete. Sections serialize as FIELD: lines so QwenVL routes every part into the right output section of the structured MiniMax prompt.

Changed

• Style dropdown options now inject STYLE: __mmh3/style/<name>__ — the resolved style lands as an explicit STYLE: field line in the prompt (deterministic routing) while keeping its [TAG] directive marker.
• Bundled mmh3/vfx wildcards synced from Garage (10 categories: energy, explosion, sparks, smoke, dust, debris, liquid, rift, lightning, feedback).

Full Changelog: v0.1.10...v0.1.11

v0.1.10

What's changed

Changed

• Style dropdown is now wildcard-driven — every option injects __mmh3/style/<name>__ instead of a hardcoded sentence, so entries stay in sync with Garage and pick a random phrasing per job; the dropdown shows the [TAG] label only (full value on hover).
• Added [MATCH REF1] / [MATCH REF2] options to style-match a single reference picture (Picture 1 or 2) — other references contribute content only, not style.

Fixed

• Synced bundled wildcards with Garage: [MUSIC] directive prefixes on all mmh3/music entries (routes music into non_diegetic_music), [TAG] directive prefixes on mmh3/style, matchref split into matchref1/matchref2, pmp drift.

Full Changelog: v0.1.9...v0.1.10

v0.1.9

Full Changelog: v0.1.8...v0.1.9

v0.1.8

Full Changelog: v0.1.7...v0.1.8

v0.1.7

New recipe: PimpMyPony Face Detailer — Pony/Illustrious T2I with a YOLO+SAM face-detail pass (Easy-Use detailerFix) on the qwen21 endpoint. Detailer denoise/steps and face-detect threshold exposed.

Recipes

• pmpIncaseStyle checkpoint added to both Pony recipe option lists
• Wildcard preset gains None — expand wildcards with no model prefix/negative injected

Assets

• macOS (Apple Silicon): ForgeHub-0.1.7-arm64.dmg
• Windows: ForgeHub.Setup.0.1.7.exe
• Linux: .AppImage / .deb

v0.1.6

Point all MiniMax H3 recipes at the hub-deployed endpoint (s0iopnieywyu7k).

Chat

• System prompt is now recipe-aware: on mmh3-* recipes the assistant knows the MiniMax spec-prompt format (subject_definitions, <Picture N>/<Subject N> ordering, [Shot N] timestamps, [D] dialogue tags, soundscape) so it can write ready-to-paste prompts

Speed

• QwenVL enhancer nodes: keep_model_loaded + max_tokens 2048 — the ~8 GB GGUF stays resident between jobs instead of reloading every run (~2 min saved on every job after the first)

Recipes

• PimpMyPony: added None to the wildcard preset options — expand wildcards with no model prefix/negative injected

Fixed upstream (worker v1.0.7)

• Config presets (steps/sampler/scheduler/sigma shift) never applied — every job ran template defaults. Turbo now correctly runs 8 steps euler+simple with shift 6/3

Assets

• macOS (Apple Silicon): ForgeHub-0.1.6-arm64.dmg
• Windows: ForgeHub.Setup.0.1.6.exe
• Linux: .AppImage / .deb

v0.1.5

ForgeHub v0.1.5 — Singularity

Desktop app for running ComfyUI workflows on Runpod Serverless — bundled backend, wildcard prompts, endpoint auto-detect, output gallery with reference reuse.

Highlights in this build

• MiniMax H3 Singularity R2VA — new recipe on the fused Singularity checkpoint: up to 3 reference images + a reference video, native and turbo presets (r2va_singularity / r2va_singularity_turbo)
• Style LoRA selector — none / realism / char_swap on Singularity recipes
• Real video upload — reference video is now a true file picker (up to 256MB), shipped base64 to the worker
• vid/ wildcard chips — quick-insert bar on every MiniMax prompt field, __vid/act/*__ and more
• R2VA fixes — single reference image works (empty loaders dropped worker-side); ModelPreviewOverrideKJ bypassed (taeh3 not required)

Downloads

Platform	File
macOS (Apple Silicon)	ForgeHub-0.1.5-arm64.dmg
Windows	ForgeHub.Setup.0.1.5.exe
Linux	ForgeHub-0.1.5.AppImage · forgehub-app_0.1.5_amd64.deb

macOS: the app is unsigned — after installing, run xattr -cr /Applications/ForgeHub.app (or right-click → Open) to bypass Gatekeeper.

Endpoint: pair with the runpod-minimax-h3 worker v1.0.5+ for the Singularity stack.

v0.1.4

ForgeHub v0.1.4

Desktop app for running ComfyUI workflows on Runpod Serverless — bundled backend, wildcard prompts, endpoint auto-detect, output gallery with reference reuse.

Highlights in this build

• Serverless-only — local/direct ComfyUI execution removed; every workflow is pinned to its own Runpod endpoint (MiniMax → comfyui-mmh3, Qwen → comfyui-qwen21), no misrouting regardless of the Settings selection
• MiniMax LoRA wiring — R2VA/T2VA/FL2VA API workflows now include LoraLoaderModelOnly: r2va_native loads the mandatory ref-LoRA, turbo presets load their turbo-LoRAs
• MiniMax R2VA fix — VHS_VideoCombine missing-node error resolved (image now ships libgl1/libglib2.0-0 for VideoHelperSuite/OpenCV)
• Enhancer presets per family — each workflow dropdown exposes only its own presets (I2VA/R2VA/FL2VA/T2V, SFW+NSFW, FL2VA incl. Loop variants); R2VA/FL2VA presets now preserve [D]…[/D] spoken dialogue verbatim and never invent speech (fixes gibberish/non-English audio). VID › Summary added to Qwen Edit.
• Media fixes — output/input previews and lightbox work in the packaged app (absolute backend URLs); new "Use as reference" button in the lightbox
• macOS 26 click fix, badge/README refresh

Downloads

Platform	File
macOS (Apple Silicon)	ForgeHub-0.1.4-arm64.dmg
Windows	ForgeHub.Setup.0.1.4.exe
Linux	ForgeHub-0.1.4.AppImage · forgehub-app_0.1.4_amd64.deb

macOS: the app is unsigned — after installing, run xattr -cr /Applications/ForgeHub.app (or right-click → Open) to bypass Gatekeeper.