Skip to content

v1.6.0 — October ecosystem sync: MOSS-Image / FLUX 3 Klein

Choose a tag to compare

View all tags
@IvenKooLab IvenKooLab released this 30 Oct 16:21
· 16 commits to main since this release
v1.6.0
efb9b9b

Changelog

  • MOSS-Image text-to-image (built in) (OpenMOSS, Apache-2.0): graph rebuilt to match the official template — SFT int8 UNET + qwen3.5-9b w4a8 text encoder (CLIPLoader type=moss_image) + flux2-vae_small; 25 steps er_sde / cfg 2.5 with layer-22 skip-layer guidance. Strong bilingual long-text rendering and layout control
  • FLUX 3 Image Klein text-to-image (built in) (Black Forest Labs): distilled int8, 16-step euler/simple (swap in the base weights for 50 steps / cfg 3.5); native 4K, text rendering, bounding-box layout, multi-reference editing
  • Two new model suites: the MOSS-Image files (plus edit and color-depth-control LoRAs) download in one click via hf-mirror; FLUX 3 Klein is listed with honest gating notes — the BFL repo requires signing in to Hugging Face and accepting the license, while the shared VAE still downloads in one click
  • Text override fixed (critical): prompt overrides previously leaked into every node holding a text input, silently overwriting the negative prompt on dual-CLIPTextEncode graphs; only the first text node is written now, and H3 single-text-node workflows are unchanged
  • The Create page ETA hint now follows the selected workflow instead of a hardcoded Flux / 20 steps, and refreshes on mode switch
  • Built-in help updated (EN + 中文); roadmap synced with the October ecosystem (NVIDIA VFX nodes, H3 single-head LoRA, the official Comfy Agent naming overlap)
  • Test portability fixed: test_models.py / test_upload_media.py hardcoded a developer-machine path and crashed everywhere else
  • 13 new regression tests for the new built-in graphs, model suites and the text-override scoping

Requires ComfyUI ≥ 0.42 for the two new built-in workflows. Full changelog: CHANGELOG.md. Requires local ComfyUI and the WebView2 runtime.

更新日志

  • MOSS-Image 文生图（内置）（OpenMOSS · Apache-2.0）：按官方模板还原——SFT int8 主模型 + qwen3.5-9b w4a8 编码器（CLIPLoader type=moss_image）+ flux2-vae_small；25 步 er_sde / cfg 2.5，跳层 22 细节锚定。中英长文本渲染与排版控制强
  • FLUX 3 Image Klein 文生图（内置）（Black Forest Labs）：蒸馏 int8，16 步 euler/simple（可在编辑器换 base 权重走 50 步 / cfg 3.5）；原生 4K、文字渲染、边界框排版、多参考图编辑
  • 新增两个模型套件：MOSS-Image 全套（含 edit 与色深控制 LoRA）hf-mirror 一键断点续传；FLUX 3 Klein 如实标注 gated——BFL 仓库需登录 Hugging Face 接受协议后手动下载，共享 VAE 仍可一键下载
  • 提示词覆盖修复（严重）：此前会写进所有带 text 输入的节点，静默覆盖双 CLIPTextEncode 工作流的负向提示词；现在只写第一个文本节点，H3 单文本节点工作流行为不变
  • 创作页预估提示不再写死「Flux · 20 步」，改为跟随所选工作流，模式切换即刷新
  • 帮助页同步（中英双语）；路线图对齐十月生态（NVIDIA VFX 节点、H3 单头 LoRA、官方 Comfy Agent 命名重叠注记）
  • 测试可移植性修复：test_models.py / test_upload_media.py 硬编码开发机路径，在别的机器上必崩
  • 新增 13 条回归测试，覆盖新内置执行图、模型套件与提示词覆盖范围

两个新内置工作流需 ComfyUI ≥ 0.42。完整变更：CHANGELOG.md。需要本地 ComfyUI 与 WebView2 运行时。

Assets 3
Loading