🚀 终极全能工作流一体化加载器

What it replaces

Five nodes and a pile of wires. CheckpointLoaderSimple, a LoRA loader, a text encoder for the positive, another for the negative, an empty latent, plus CLIPLoader and VAELoader if your checkpoint is missing either half. UltimateAllInOneLoader folds all of it into one box with five outputs: 模型, 正向提示词, 反向提示词, 潜空间, VAE.

If that were all it did, it'd be a convenience node. The reason it's worth an article is the second thing it does: it can read a prompt file and emit one conditioning pair per prompt group, which turns a single queued workflow into a batch of generations without you touching a batch setting.

The mechanism worth understanding

正向提示词 and 反向提示词 come out as lists of CONDITIONING, not single values. In ComfyUI, connecting a list where a sampler expects one conditioning means the sampler gets executed once per entry - in this node's author's own words in the source, downstream "will sample once per group." One queue press, N generations, N different prompts. That's the whole feature, and it's easy to miss because nothing on the canvas announces it.

Each group comes from a plain text file, in a format defined by two prefixes:

positive: masterpiece, best quality, 1girl, standing on a rooftop at night
negative: low quality, worst quality, blurry, extra fingers

positive: cinematic photo of the same girl, rain, neon reflections
negative: low quality, worst quality, blurry, oversaturated

A group is committed when a negative: line follows a pending positive: line, so a stray positive with no negative after it is silently dropped. Points to be aware of: the path is honoured as an absolute path if it is one, otherwise it's resolved against ComfyUI/input, and if the file doesn't exist the node prints a line to console and falls back to masterpiece, 1girl / low quality - so a typo'd path gives you a render, just not the one you asked for. 起始索引 and 读取数量 slice the file (0 in 读取数量 means read everything), which is how you resume a 200-prompt file in chunks instead of babysitting it.

If you'd rather skip files entirely, flip 手工输入提示词 to 启用 and the two text boxes are encoded as a single group. When it's on, the file is ignored completely - worth remembering, because that's a common "why is my prompt file doing nothing."

The rest of the inputs

Checkpoint模型 and LoRA模型 are dropdowns, and the entries carry annotations the author generated by reading the safetensors header: file size, plus a tag like [3合1完整版] for a full model+CLIP+VAE bundle, [纯Unet/精简版] for a bare UNet, [无VAE/2合1], or [自定权重版]. That's useful before you load anything. 可选CLIP and 可选VAE are how the node handles a checkpoint that ships neither half, and it relaxes its own input validation because those annotated strings would otherwise break saved workflows - it strips the annotation when it actually loads the file.

CLIP类型 matters more than it looks. It's a 28-entry list of text-encoder types, and the author's tooltip gives the two cases people get wrong: Anima's qwen_3_06b uses stable_diffusion, while Qwen-Image needs qwen_image. CLIP跳过层 is A1111's clip_skip and defaults to 1 (i.e. off); anime checkpoints usually want 2. It's applied only for SD-style text encoders - the source checks the filename for qwen/llama/gemma/umt5/anima and skips the clip-skip step for those rather than corrupting the encoding, printing a note so you know it was ignored. 模型强度 and CLIP强度 are the LoRA strengths, and the LoRA is skipped entirely if both are zero.

宽度, 高度 and 批次大小 drive the 潜空间 output. One thing to keep in mind: the loader builds that latent the plain-SD way, while the sibling test node in this pack builds its empty latent from the model's latent_format (the source explicitly handles models with 16-channel 3D latents, Anima/Wan-style). So if you're on one of those architectures, don't assume this node's 潜空间 is shaped right for your sampler - use the test node's path or an explicit latent node instead.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/kukuxz007/comfyui-t2i-tester.git

Restart ComfyUI; no dependencies to install for this node (requirements.txt is only for the optional hand-repair node). In ComfyUI Manager, search the Registry for ComfyUI T2I Tester; the README notes newer Manager builds only install Registry packages, so fall back to the clone or the legacy Install via Git URL box if the search is empty. You'll find it under the 自定义工作流 category.

Two things to know before you rely on it

The node ships a file-picker for Prompt文件, backed by a server endpoint that lists directories and .txt files anywhere on the machine - it'll happily read an absolute path outside ComfyUI/input. Convenient; just remember ComfyUI has no authentication by default, so don't expose that port directly to the internet.

And the list semantics cut both ways: because the conditioning outputs are lists, anything you wire them into downstream that expects exactly one prompt gets executed once per group - which is the point, but it also means a 30-group file queues 30 full generations with one click. Test with 读取数量 set to 1 before you leave the room.