图片｜GPT-Image-2 官方修图 V2

This is the pack's flagship image editor, and it's a different beast from the legacy -C edit node. TikpanGptImage2OfficialEditV2 hits Tikpan's multipart /v1/images/edits endpoint with the full modern parameter set: a required main image, an optional mask for region-restricted redraws, up to 15 reference images (plus a 参考流 IMAGE-batch input), resolution/aspect/size control, quality tiers, background mode, and a batch mode that submits multiple edit requests concurrently. If the pack has a "do serious image editing in ComfyUI" node, this is it.

How it works

The 主图像 is uploaded via multipart along with any reference images and mask; the node locally preprocesses the image to the chosen size tier (分辨率档位) and aspect ratio before submitting, then polls for the result. The 生成张数 field (1–10) is handled smartly: instead of one giant request, the node submits one request at a time (n=1 each) to dodge per-request limits, optionally concurrent (并发请求数, up to 4). Local knobs like 提示增强 (appends quality phrasing to your edit instruction - not an upstream API param), 遮罩反相, and 超时秒数 round out the batch-control layer. Outputs: 编辑结果 (IMAGE) and 渲染日志 (STRING).

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 主图像 - required; the image being edited.
• 编辑指令 - what to change (default is a well-written brief: edit per the requirements, and if a mask is provided, only modify the masked region, keep the rest unchanged - worth reading, it's the model of a good edit prompt).
• 遮罩掩码 - optional; white = redraw, black = keep.
• 参考图1 … 参考图15 / 参考流 - up to 15 references, or a whole IMAGE batch in one wire.
• 分辨率档位 / 画面比例 / 尺寸 - three stacked controls; 尺寸 overrides the first two when set, otherwise Auto.
• 画质 - auto / low / medium (default) / high. Cost scales up with quality.
• 生成张数 and 并发请求数 - batch volume and how fast it burns through it.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart (or Manager → "Tikpan"). No models - but note the 请求 scale: concurrency × generations is how a single node run multiplies spend.

Where people get burned

• Local params aren't API params. 提示增强, 遮罩反相, 并发请求数, and 超时秒数 are explicitly local pre/post-processing knobs - the schema's tooltips say they're not passed upstream. Expect results consistent with that (e.g., mask invert happens before upload).
• Concurrency is a spend multiplier. 4 concurrent × 10 generations = 40 billed edits racing. Start at 2 and check latency before cranking it.
• Mask polarity. White redraws, black preserves - same convention as the legacy editor, and inverted masks are the classic silent failure.
• Background mode is reduced. The background param exposes auto/opaque; the tooltip notes transparent isn't surfaced and old workflows get treated as auto. If you need true transparent backgrounds, that path isn't in this UI.
• Size overrides can surprise. Setting a hard 尺寸 beats the aspect-ratio setting - know which knob is on top.

The bottom line: this is the editor to use for real GPT-Image-2 edit work - masked inpainting, multi-reference fusion, and batch output in one node. Read the tooltips on the local-vs-upstream params once and it's genuinely powerful; skip them and the batch controls will cost you.