nodes/Xyc2016-Comfyui_Fd_Nodes
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
214 Commits 28 Branches 169 Tags
bc5a032164d4989ecf8dc633dfd217873d1d6820
Commit Graph
214 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
xuyucai bc5a032164 test: provide a tensor to watermark upload guard test 2026-10-29 11:31:01 +08:00
xuyucai ce12ca45ac test: update overseas watermark guard for URL uploads 2026-10-29 11:30:05 +08:00
xuyucai e3cadf434a feat: support watermark URL delivery for overseas nodes v53.8.71 2026-10-29 11:26:45 +08:00
Modesto Schmeler f01eb85ca0 feat: release overseas URL image delivery v53.8.70 
feat: overseas URL image delivery v53.8.70
v53.8.70 		2026-10-28 21:11:33 +08:00
chenjunhong 1125eda802 feat: enforce overseas URL image delivery in v53.8.70 2026-10-28 21:08:24 +08:00
chenjunhong bc62add2ce feat: support overseas GPU endpoints and OSS URL image delivery in v53.8.69 v53.8.69 2026-10-28 18:29:37 +08:00
xuyucai 048601bc64 chore: bump version to 53.8.68 for qwen detection retries v53.8.68 2026-10-28 15:04:43 +08:00
xuyucai 1654ee11a1 fix: retry transient failures in qwen detection 2026-10-28 14:50:07 +08:00
chenjunhong be948c1cf1 chore: bump version to 53.8.67 for image-generation image-to-text routing v53.8.67 2026-10-23 18:31:11 +08:00
chenjunhong e79426721f feat(doubao): route image-to-text via image-generation when FD_DOUBAO_VISION_URL set 
- single-image node: image_base64 to /image/image_to_text_doubao
- combo node: image_base64_list for multi-image slots, concurrent path kept
- server side adapts CN/overseas model naming (dot vs dash), caller agnostic
- unset URL keeps LiteLLM direct path unchanged for CN deployments
 2026-10-23 18:20:33 +08:00
chenjunhong 26fe408f69 feat(doubao): accept image-generation image_to_text_doubao response 
FD_imgToText_Doubao now parses both the legacy gateway response
({response: {Result}}) and the image-generation /image/image_to_text_doubao
response ({status, result_text}), so overseas deployments can point
FD_DOUBAO_URL at the cluster-internal image-generation service without
code changes. Also skip the Authorization header when FD_DOUBAO_KEY is
unset (previously raised TypeError and silently returned defaultPrompt);
the image-generation endpoint needs no auth from inside the cluster.
 2026-10-23 13:56:15 +08:00
chenjunhong d47485ddfd fix: correct 21:9@4K pixel pin and reject unsupported gemini ratios in v53.8.66 
- gpt-image-2 4K+21:9 no longer reuses 16:9's 3840x2160: the gateway
  passes explicit pixel sizes through verbatim and ignores ratio, which
  produced 16:9 output for 21:9 requests. Pin to 3840x1648 (matches
  resolution_to_edit_size and gateway ASPECT_RATIO_SIZE_TABLE)
- Gemini nodes no longer offer 9:21 (upstream/Gateway whitelist does not
  support it; it was silently dropped, making output follow the first
  input image). Applies to ZhiYiImageToImageNode, ZhiYiImageToImageComboNode,
  FD_GeminiImage; aistudio test node keeps it for gpt/aistudio_publish paths
- Add validate_gemini_aspect_ratio guard (mirrors gateway whitelist) at
  build_request_body, litellm client and old gemini node entry points so
  unsupported ratios fail loudly instead of silently
- Contract tests: node size output must be tier shorthand or ratio-exact
  (<0.5%) 16-aligned pixels; gemini nodes must only offer supported ratios
- Bump version to 53.8.66
v53.8.66 		2026-10-22 17:30:29 +08:00
chenjunhong d06c79c202 chore: drop duplicated OkResponse.content assignment left by rebase 2026-10-19 18:34:17 +08:00
chenjunhong f1d28f95ca fix: restore missing 4:5/5:4 seedream ratios and align size map with server in v53.8.65 
- Add 4:5/5:4 to SEEDREAM_SIZE_MAP for all tiers (previously fell back
  to 1:1 silently even though exposed in the aspect ratio enum)
- Align 2K/4K pixel values with model-servings ASPECT_RATIO_SIZE_TABLE
  (Seedream 5.0), interpolate missing ratios for the 3K tier
- Warn instead of silently falling back on unknown ratios
- Guard test: every exposed ratio must exist in the size map
- Fix stale test stubs (outdated OSS assertions, wrong tuple index in
  combo 3K test); add flash 4:5/5:4 regression cases
- Bump version to 53.8.65 (v53.8.64 already released upstream)
v53.8.65 		2026-10-19 18:23:36 +08:00
xuyucai bece2d00d6 Release 53.8.64 with restored OSS download regression coverage v53.8.64 2026-10-16 10:45:35 +08:00
xuyucai 5eb6f5e2de Use internal Hangzhou OSS URLs for node image downloads 2026-10-16 10:42:07 +08:00
chenjunhong b7cdc75338 fix: prevent four-grid background subject leakage in v53.8.63 v53.8.63 2026-10-14 17:57:00 +08:00
chenjunhong 8940be46f5 feat: add nano banana 2.1 explicit channel models in v53.8.61 v53.8.61 2026-10-14 11:06:49 +08:00
chenjunhong b7bb8f6a20 feat: upgrade four-grid crop backgrounds and aspect ratio in v53.8.60 v53.8.60 2026-10-12 17:24:24 +08:00
chenjunhong 0bc3939dae feat: add nano banana 2.1 and four-grid crop nodes v53.8.59 2026-10-10 16:37:29 +08:00
chenjunhong 1c43916edc feat: switch qwen detect default model from qwen3-vl-flash to qwen3.6-flash in v53.8.58 v53.8.58 2026-10-09 15:02:44 +08:00
chenjunhong 9cbd0d4281 chore: bump version to 53.8.57 v53.8.57 2026-09-24 17:20:05 +08:00
chenjunhong 5939c31577 feat: add gpt-image-2.5-sunburst/flare alias channels in v53.8.56 v53.8.56 2026-09-24 17:06:56 +08:00
chenjunhong 7699c20f4b feat: add doubao-seedream-5.0-flash model option in v53.8.55 v53.8.55 2026-09-23 16:10:51 +08:00
chenjunhong fffefcea95 feat: support gpt-image-2.5-vip channel and bump timeout to 600s in v53.8.54 v53.8.54 2026-09-23 13:53:59 +08:00
chenjunhong 69d7da72f7 chore: bump version to 53.8.53 v53.8.53 2026-09-10 11:00:32 +08:00
chenjunhong 920a385bb4 feat: support GPT Image 2 primary and increase timeout in v53.8.52 v53.8.52 2026-09-10 10:28:19 +08:00
chenjunhong 62e44c5df9 fix: route all GPT 2.5 channels through image-generation service v53.8.51 2026-09-09 18:36:12 +08:00
chenjunhong 651fd3ed93 feat: add GPT 2.5 sunburst and flare LiteLLM model options v53.8.50 2026-09-09 18:12:09 +08:00
chenjunhong 9d43565cb9 feat: support GPT Image 2.5 and five quality levels across GPT nodes v53.8.49 2026-09-09 17:44:06 +08:00
chenjunhong 54a1e6e183 feat: support five image quality levels and bump version to 53.8.48 v53.8.48 2026-09-09 16:54:54 +08:00
chenjunhong a148e77bcc chore: align package version to 53.8.47 v53.8.47 2026-09-01 15:42:55 +08:00
chenjunhong 8b9c4f71fc fix: make RMBG image output follow background setting v53.8.46 2026-09-01 15:23:07 +08:00
chenjunhong 25cdfb9b50 feat: add direct RGBA output for RMBG nodes v53.8.45 2026-09-01 11:12:55 +08:00
chenjunhong f0db33bf6b chore: bump version to 53.8.44 v53.8.44 2026-08-21 19:39:27 +08:00
chenjunhong b44b3bc89b fix: normalize 3-dim image tensors before downscale in seedream upload 2026-08-21 19:39:20 +08:00
chenjunhong 52118e16e8 chore: bump version to 53.8.43 v53.8.43 2026-08-21 19:23:09 +08:00
chenjunhong 7ecce0992f feat: route Seedream nodes through image-server with litellm fallback 2026-08-21 19:23:06 +08:00
chenjunhong 2aea2fe1f4 feat: support Gemini 9:21 aspect ratio v53.8.42 2026-08-20 18:16:49 +08:00
chenjunhong 2cdda6d928 chore: bump version to 53.8.41 v53.8.41 2026-08-17 17:06:21 +08:00
chenjunhong 90ca2c4a0a feat: extend pattern extraction nodes with alpha apply node 2026-08-17 17:05:17 +08:00
chenjunhong 2e6969ef75 chore: bump version to 53.8.40 v53.8.40 2026-08-11 15:23:26 +08:00
chenjunhong 7c31b796df fix: switch class widgets to new-style COMBO so multi-select passes validation 
ComfyUI 0.6.0 后端 combo 校验（execution.py）只对"input_type 是选项列表"的老式
写法做成员判断；此前 classes 用老式 (选项列表, {...}) 写法，多选下拉提交的
列表值必被 "Value not in list" 拒绝。改为新式 ("COMBO", {"options": [...],
"multi_select": {...}}) 写法：
- input_type 是字符串 "COMBO"（非列表），后端解包前端 {"__value__": [...]}
  包装的数组后跳过 combo 校验，列表值直接传给节点
- 前端 1.34.9 通过 isComboInputSpecV2 识别新式 COMBO，渲染多选下拉
多选下拉恢复可用；中文显示需前端 label 映射（当前版本多选对对象 options
原样提交，翻译表保留作参考）。
 2026-08-11 15:23:00 +08:00
chenjunhong b5c3ea8a76 chore: bump version to 53.8.39 v53.8.39 2026-08-11 14:40:49 +08:00
chenjunhong 64625c4346 fix: class multi-select options submitted as objects fail validation 
ComfyUI 0.6.0 前端 MultiSelectWidget 会把选中的整个 {text, value} 对象作为值提交，
后端 prompt 校验做成员判断时对象数组永远不匹配 options，导致 "Value not in list"
校验失败。将衣物/时尚/身体 classes 下拉改为纯英文字符串 options
（list(*_CLASSES)），前端提交 ["Face", ...] 英文数组、后端逐元素校验通过。
中文标签保留翻译表供后续前端扩展使用。
 2026-08-11 14:40:19 +08:00
chenjunhong 45bc4b0f9f chore: bump version to 53.8.38 v53.8.38 2026-08-11 10:05:42 +08:00
chenjunhong e3585f9153 fix: env dropdown turns into object after switching 
env 单选 Combo 不支持 {text, value} 对象 options（那是多选 widget 的机制），
选中后整个对象被序列化为值导致 node 报错。修复：
- env 下拉改回纯中文字符串（灰度环境/开发环境/自定义（使用 service_url））
- 后端 normalize_env 归一化中文文本为 gray/dev/custom，并兼容旧工作流
  已损坏的 {text, value} 对象与旧 key 字符串
 2026-08-11 10:05:36 +08:00
chenjunhong 6dd1b1dad6 chore: bump version to 53.8.37 v53.8.37 2026-08-11 09:49:29 +08:00
chenjunhong 01861ab91b feat: add 8 v2 external API nodes with env switch and class multi-select 
新增 8 个 v2 节点（RMBG/衣物/时尚/身体分割、DWPose、LineArt、Depth、SAM2），
解决环境切换靠手填 URL、分割类别手打英文、服务端开关硬编码三个痛点：
- env 下拉切换灰度/开发/自定义环境，gray/dev 使用内置 base url 忽略 FD_* 环境变量
- 衣物 18/时尚 47/身体 12 类别多选下拉（中文显示、英文提交），multiselect + multi_select 标记
- DWPose return_pose_image/return_openpose_json、LineArt return_image 开关，关闭时原图直通
- SAM2 补齐 node_switch 短路（batch mask 与输入一致），发请求前校验 background/background_color

老节点零改动：基类仅加默认参数与缺失容错，老调用路径行为不变。
 2026-08-11 09:49:22 +08:00
chenjunhong f5df28120c chore: bump version to 53.8.36 v53.8.36 2026-08-06 14:53:17 +08:00