🎨 AniSee — Official ComfyUI Workflow

Step-by-Step Layout • Auto Quality Prefix • Optional Prompt Enhancer • LoRA + Upscaler Ready

Built for AniSee Base • Drop-in for Anima Preview3 • 1 MP Native

✨ What is this Workflow?

This is the official ComfyUI workflow for AniSee — my personal anime fine-tune of Anima Preview3 Base. It is designed for one thing: get clean anime results out of AniSee with as little prompt-writing effort as possible, while still giving full control over every stage.

The workflow is laid out as a clear left-to-right pipeline with numbered step groups (📦 Load → 📖 LoRA → ✍️ Write → 🎲 Enhance → 🎨 Generate → 🔼 Upscale), so you can see at a glance where everything happens.

It is the foundation workflow for the AniSee family — future workflows for the planned AIO and 4-Step Turbo variants will follow the same structure with adapted nodes.

🎯 Key Features

• ✅ Automatic quality prefix — no need to type masterpiece, best quality, score_7… yourself

• ✅ Optional Qwen3-VL prompt enhancer — turns one-line ideas into full Danbooru tag lists

• ✅ LoRA support via Lora Manager (bypassed by default, toggle on with a single click)

• ✅ UltimateSDUpscale for 2× upscaling (bypassed by default)

• ✅ Image comparer built in — see base vs upscaled side by side

• ✅ Toggle bypass groups — flip whole sections on/off with one click

• ✅ Pre-configured with my preferred settings: er_sde + simple + 40 steps + CFG 4.5

• ✅ Pre-loaded recommended negative prompt

🖼️ Example Gallery

See what AniSee can produce with this workflow:

🎨 View the AniSee Sample Gallery →

📁 Required Model Files

ComfyUI/models/diffusion_models/
└── AniSee.safetensors
ComfyUI/models/text_encoders/
└── qwen_3_06b_base.safetensors
ComfyUI/models/vae/
└── qwen_image_vae.safetensors
ComfyUI/models/upscale_models/        (only if using the upscaler)
└── 4x-UltraSharp.pth

If you already run Anima Preview3 Base, you already have the text encoder and VAE — AniSee is a direct drop-in.

The upscale model 4x-UltraSharp.pth can be downloaded here:

• OpenModelDB — 4x-UltraSharp

• HuggingFace — Kim2091/UltraSharp

Any other 4× ESRGAN-style model will also work — just swap the file name in the 📦 Load Upscale Model node.

🗺️ Workflow Structure

The workflow is organized into six numbered step groups:

StepGroupPurpose📦 Step 1Load AniSee & ModelsUNet, Text Encoder, VAE📖 Step 2Load LoRA (Optional)Stack LoRAs via Lora Manager✍️ Step 3Write PromptType your subject — quality tags added automatically🎲 Step 4Prompt Enhancement (Optional)Qwen3-VL converts ideas to full tag lists🎨 Step 5Generate ImageKSampler + Decode + Save🔼 Step 6Upscale (Optional)UltimateSDUpscale 2× + side-by-side compare

🔗 The Quality Prefix Node

One of the central conveniences of this workflow is the 🔗 Quality Prefix node (a StringConcatenate). It sits between your prompt input and the CLIP Text Encode and automatically prepends:

masterpiece, best quality, score_7, highres, illustration,

to every prompt before it reaches the encoder. You don’t need to type quality tags yourself — just write the subject and scene you actually care about:

1girl, long silver hair, blue eyes, rainy city street at night, neon lights, cinematic lighting

…and the prefix turns that into:

masterpiece, best quality, score_7, highres, illustration, 1girl, long silver hair,
blue eyes, rainy city street at night, neon lights, cinematic lighting

If you want to experiment with different quality tag combinations, just edit the prefix string directly in the node.

🎲 The Prompt Enhancement Group

The 🎲 Step 4 - Prompt Enhancement group is the magic shortcut of this workflow. It is a small subgraph built around Qwen3-VL-4B-Instruct (GGUF) that acts as an automatic prompt-to-tags converter.

When enabled, you can write a prompt as short as a single phrase, and it will be expanded into a fully-formed Danbooru-style tag list with proper tag ordering: subject → face → hair → eyes → pose → clothing → background → lighting → style.

How it works in practice

Your one-liner input:

cute anime girl in a garden

Becomes after the enhancer:

1girl, solo, cute, smile, looking_at_viewer, long_hair, brown_hair, brown_eyes,
sundress, white_dress, standing, garden, flowers, sunlight, depth_of_field

This output then flows into the same 🔗 Quality Prefix → CLIP Text Encode chain as a regular prompt — so the quality tags still get auto-prepended.

Combined with the Quality Prefix

Together, the Prompt Enhancement group and the Quality Prefix node mean you literally only have to type the idea of an image. Everything else — quality tags, tag formatting, tag ordering — is handled automatically. Great for fast brainstorming or batch generation.

Enabling / Disabling

By default the Prompt Enhancement group is bypassed. To turn it on, right-click the group or the 🎲 Prompt Enhancement node and toggle bypass off. The Qwen3-VL GGUF model will need to be available to ComfyUI-QwenVL on first use (the node will download or expect it depending on your setup).

📖 LoRA Section

The 📖 Step 2 - Load LoRA group lets you stack one or more LoRAs onto AniSee. It is built around Lora Manager for easy selection and trigger word handling. By default the group is bypassed — toggle it on via the 📖 Toggle LoRA ON/OFF button.

The 🏷️ LoRA Trigger Words node handles trigger word injection automatically, so you don’t need to memorize which tags belong to which LoRA.

🔼 Upscale Section

The 🔼 Step 6 - Upscale group uses UltimateSDUpscale with chessboard tiling to bring 1 MP base images up to 2× resolution.

Default upscaler settings:

• Upscale by: 2×

• Steps: 20

• CFG: 4.5

• Sampler / Scheduler: er_sde / simple (matches base)

• Denoise: 0.22 — keep it low to preserve faces and composition

• Tile size: 1024 × 1024, mask blur 16, tile padding 64

Toggle the upscaler on with the 🔼 Toggle Upscaler ON/OFF button. The built-in 🔍 Compare Images node lets you A/B the base and upscaled versions side by side.

⚙️ Default Sampler Settings

Steps:       40
CFG:         4.5
Sampler:     er_sde
Scheduler:   simple
Resolution:  832 × 1216  (portrait — change in 📐 Image Size)

Good resolutions to try: 1024×1024, 896×1152, 1152×896, 1254×836, 1365×768. Stay around 1 MP for the cleanest results.

➖ Default Negative Prompt

The negative prompt is pre-loaded — no need to type anything:

worst quality, low quality, score_1, score_2, score_3, artist name,
(lowres:1.2), (worst quality:1.4), (low quality:1.4), (bad anatomy:1.4),
bad hands, multiple views, comic, jpeg artifacts, patreon logo,
patreon username, web address, signature, watermark, artist name,
censored, mosaic censoring

Feel free to soften the weights if your images come out too flat.

📦 Required Custom Nodes

All of these are installable via ComfyUI Manager:

• ComfyUI-Easy-Use — by yolain

• ComfyUI_UltimateSDUpscale — by ssitu

• ComfyUI-Lora-Manager — by Will Miao

• ComfyUI-QwenVL — by AILab (1038lab)

• rgthree-comfy — by rgthree (for Seed, Image Comparer, Fast Groups Bypasser)

🚀 Quick Start

1. Install all required Custom Nodes (via ComfyUI Manager).

2. Place AniSee, the Qwen text encoder and the Qwen-Image VAE in the correct folders.

3. Drag this workflow JSON onto ComfyUI.

4. Type a short prompt into the ✨ Positive Prompt node — quality tags are auto-added.

5. (Optional) toggle on the 🎲 Prompt Enhancement group for fully automatic Danbooru tagging.

6. Hit Queue.

📈 Version History

v1.0 — Initial Release

• Six-step pipeline layout: Load → LoRA → Write → Enhance → Generate → Upscale

• Auto Quality Prefix via StringConcatenate

• Optional Qwen3-VL-4B prompt enhancement subgraph

• Optional LoRA stack via Lora Manager

• Optional UltimateSDUpscale 2× with side-by-side compare

• Pre-configured with recommended settings and negative prompt

🙏 Credits

• AniSee Model: SeeSee21 — based on Anima Preview3 Base by CircleStone Labs and Comfy Org

• Custom Nodes: yolain, ssitu, Will Miao, AILab (1038lab), rgthree

• Upscale Model (default): 4x-UltraSharp by Kim2091

AniSee Workflow — clean anime generation, minimum prompting effort. 🎨