1,500 Anima artist styles on RunningHub for about ~1-5 coin each

I wanted to see how far Anima's artist tags really go. That meant thousands of test pictures, too slow for one local GPU, so I sent them to RunningHub.

The recipe

Setting	Value
Model	anima-base-202605-v1.0.safetensors (RunningHub's name for Anima base v1.0)
Text encoder / VAE	qwen_3_06b_base / qwen_image_vae
Speed LoRA	Turbo-ANIMA-v4 from RunningHub's public library
Size	768x1344
Steps / CFG	8 / 1
Sampler / scheduler	er_sde / simple
Tier	Standard

Same prompt for every picture; only the artist tag changes.

Try it yourself: download the workflow

Download the workflow (.json) Run it on RunningHub

No GPU? The RunningHub version runs in the cloud: open it as a simple app (type a prompt, press run) or launch the full workflow. About 1 RH coin per picture.

Running locally? Drag the file into ComfyUI. Then change only the @artist tag in the prompt and queue again to compare styles.

Not sure which artist to try? Two places, both work: Anima Style Gallery shows 42,000+ artists rendered with the same test prompt (right-click a card to copy its tag), and danbooru lets you look up any artist and use their tag with @ in front.

You need	Folder
anima-base-v1.0.safetensors	models/diffusion_models
qwen_3_06b_base.safetensors	models/text_encoders
qwen_image_vae.safetensors	models/vae
Turbo-ANIMA LoRA	models/loras
Anima Style Explorer node (optional)	install through ComfyUI-Manager

The prompt node is Anima Style Explorer. Without that pack, replace it with a normal CLIP Text Encode: the output is identical.

This picture has the workflow inside it too: drag it into ComfyUI to load the same graph. Rendered locally on an RTX 5080 in about 4 seconds.

Where the artists come from: Anima Style Explorer

The artist list for this test came from Anima Style Explorer (fulletLab/comfyui-anima-style-nodes), a free ComfyUI node pack.

• It replaces CLIP Text Encode. Write your prompt, open its visual gallery of artist styles, press Apply Style, and it adds the @artist tag and sends the conditioning straight to the sampler.
• 20,000 artists bundled, each with a work count. I only used artists with at least 300 works: with fewer, the style tends to be weak.
• Auto Cycle keeps queuing renders while it rotates artists, the same idea as this test but inside ComfyUI.
• Works offline. Remote preview images are opt-in.

Style explorer concept by ThetaCursed.

No node? Use the web gallery or danbooru

Anima Style Gallery by Mooshieblob does the same job in your browser, free:

• 42,189 artists on Anima 1.0, every one rendered with the same test prompt, so you compare styles and nothing else.
• Search by tag, sort by number of works or trending, hide artists with few works, or hit Random.
• Right-click a card to copy the tag straight into your prompt. Gen Params shows the settings the gallery used.

danbooru is where Anima's artist tags come from. Search an artist, check their work, and write their tag with @ in front (@huyandere). The post count next to each artist matters: artists with 300+ works give the clearest styles.

Use the gallery to see styles side by side, and danbooru to check an artist or find one the gallery does not show. Both are the easy way to pick artists when you run the workflow somewhere the Explorer node is not installed, like RunningHub.

Batching

• 25 artists per task: one SaveImage branch per artist, each with its own filename prefix so the results can be matched (RunningHub returns them out of order).
• 3 tasks at a time.
• One task took about 125 seconds for 25 pictures.

What it cost

Batch	Pictures	Coins	Per picture
1	500	460	0.92
2	500	505	1.01
3	500	529	1.06
Total	1,500	~1,494	~1.0

About 4.5 GPU seconds per picture, and roughly 20 minutes for each 500.

The surprise: styles saturate

I sorted all 1,500 pictures by visual similarity (CLIP ViT-H embeddings) against 530 artists I had already tested locally:

• 61% landed in one look: glossy, pretty, modern anime.
• Only 24 of 1,500 looked unlike anything in the first 530.

So past a few hundred artists, more tags mostly give you the same picture again. If you are hunting for distinct styles, test a few hundred well-known artists rather than thousands.

RunningHub replaces seeds, so these are not same-seed with the local tests. Read the real seed from each PNG's metadata if you want to reproduce one.