凤希AI - 千问2.1三视图生成器

Turnaround sheets used to be a project: CharTurnerV2, then Flux Kontext plus a dedicated LoRA, then manually assembling the grid. By 2026 it's one prompt against an edit model - the KB's character-consistency essay calls that substitution out, and the Qwen-Image-Edit line is where the community does it, with tooling like Lazy Character Sheet and VNCCS built on the editor rather than on adapters. (The display name, 凤希AI - 千问2.1三视图生成器, means "Qwen 2.1 three-view generator".)

What makes it different from the plain editing node it sits next to is one thing: it writes the layout half of your prompt for you.

Your prompt is a description, not an instruction

Under the hood it prepends a fixed Chinese template to whatever you type in 正向提示词. The template asks for a full-body character reference sheet built from the facial features of <image1>, pure white background, three views in a fixed order - 正面全身 (front full body), 左侧面全身 (left profile), 背面全身 (back) - no cropping at head or feet, consistent bone structure, face, hair and clothing across all three, then a final line: 用户要求 ("user requirements:") followed by your text.

So you write the character, not the spec. Body proportions, hair, outfit, art style. If you dump "generate a three view sheet of…" into there you're just repeating the template's own job, and the model splits its attention between two conflicting instructions. The author's sample prompt is a good model: it spells out roughly seven-head proportions, shoulder and hip ratios, a natural standing pose and an explicit ban on stretched legs.

Everything else is the edit node in this pack with a shorter function body: references in 图片列表 are resized to about 宽度 × 宽度 of area (aspect preserved), VAE-encoded, attached as reference_latents, and handed to the text encoder so <image1> resolves. 潜空间 is the packed 64-channel latent the 2.1 checkpoint samples, so no Empty Latent node. Leave 负向提示词 empty and you inherit the author's built-in Chinese quality negative.

Inputs and outputs

Required: clip (CLIPLoader with type qwen_image), vae (the matching 2.1 VAE), 正向提示词, 宽度, 高度. Optional: 负向提示词 and 图片列表. Outputs: 正向条件, 负向条件, 潜空间.

Both prompt fields are forceInput, so they show up as sockets with no text box - wire a multiline string node in or you'll think the node is broken. 宽度/高度 default to 960×1280, portrait, which gives three standing figures room; that default is doing more work than it looks like.

Then it's the standard tail: 正向条件 → KSampler positive, 负向条件 → KSampler negative, 潜空间 → KSampler latent, KSampler → VAEDecode → save or preview. The author's shipped graph (千问2.1三视图V1.json) runs 25 steps, euler, sgm_uniform, CFG 1, takes references from a character-image selector, and routes the decoded sheet into the pack's image manager with the folder set to views - the sheet lands in a character asset library you can reuse as reference art later. Steal that.

Install

Same install as the rest of the pack. ComfyUI Manager → search "fxai" or "凤希" → install FxAi(凤希全能节点), or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit

Restart ComfyUI afterwards. No requirements.txt; the pack pip-installs soundfile and psutil on first import, so check the console. Models you'll need: qwen_image_2.1_bf16.safetensors in models/unet, qwen_image_2.1_vae_bf16.safetensors in models/vae, and a Qwen3-VL-8B text encoder in models/text_encoders. These come from the author's own repack links (his workflows ship a note with cloud-drive URLs) - "Qwen Image 2.1" is close to invisible in English-language discussion, where the local crowd is still on 2509/2511, so don't expect a one-click HuggingFace download.

If you are on 2509/2511, this isn't your node. The pack ships a parallel older family - 三视图生成器 and 多视图生成器 - built for those checkpoints; the 2.1 pair emits the packed latent that only the 2.1 weights eat.

Common issues

The template's loudest demands exist because they're the things that go wrong, so treat them as the checklist. Feet or heads sliced off: push 高度 up (or back off to a wider sheet) and reroll - the prompt asks for head and foot room, but the model still runs out of canvas sometimes. A fourth view sneaking in, or clothing changing between panels: reroll the seed - there's nothing to tune here, it's the sampler that decides.

Feeding it a full-body shot when you want a face is the mistake that costs the most. Your reference is resized down to roughly a megapixel, so facial features survive a tight head crop far better than a small face inside a wide frame - and facial identity is the one thing the template leads with.

White background is a prompt instruction, not a guarantee. If you need clean alpha for sprite work or LoRA training data, plan a background removal pass rather than re-rolling until the model behaves.

And the caveat that applies to every Qwen-Image-Edit workflow: the model re-emits the entire frame. The sheet is close enough to be useful, not pixel-identical to its source - fine, when the sheet's job is to be the reference for everything downstream.