Anima All-in-One Workflow v5.1: i2i, Inpainting, Highres, Detailer, and Metadata Save

Recommended Civitai title: Anima All-in-One Workflow v5.1: i2i, Inpainting, Highres, Detailer, and Metadata Save

Overview

This is an English guide for the Anima All-in-One ComfyUI workflow. The workflow is built around Anima base generation and includes text-to-image, image-to-image, LLLite-based inpainting, HighRez, SAM3 detailers, USDU upscaling, LoRA style presets, NAIA prompt assistance, and Civitai-friendly metadata saving.

Version 5 focuses on making i2i and inpainting easier to use. Version 5.1 adds small fixes around Torch Compile and HighRez options, including a safer disable_dynamic_vram setting for compile-related stability.

Tested / Recommended ComfyUI Setup

• Recommended ComfyUI version: 0.20.1.
• ComfyUI versions below 0.20.0 may not support the SAM3 nodes used in this workflow.
• Some newer UI changes may break the layout of older node-based workflows, so check compatibility if you update ComfyUI.
• Recommended launch option used by the author: --disable-dynamic-vram.

The workflow also relies on several custom nodes. Make sure the Spectrum KSampler / Mod Guidance node, NAIA Bridge, DCW, Anima LLLite, rgthree, Impact Pack, KJNodes, and Image Saver nodes are installed and updated.

Optimization Stack

The workflow is designed to be used with the following optimization order:

1. Disable dynamic VRAM — use ComfyUI with --disable-dynamic-vram and keep the workflow-side disable_dynamic_vram option enabled when using Torch Compile.
2. Sage Attention — the model loader has a Sage Attention patch option. The default workflow setting uses auto with compile support enabled.
3. Anima block compile — TorchCompileModelAdvanced is used around the Anima model path. It is configured for Anima block-oriented compilation rather than compiling every unrelated part of the workflow.
4. Spectrum — Spectrum is used after the model-level optimizations, mainly around the sampler, HighRez, detailers, and USDU paths.

If you get errors, disable optimization-related features one by one first. Start with Spectrum, then Torch Compile, then HighRez or detailers. The first run after compile can also be much slower because the model is being compiled.

Important Node Groups

0. Prompt Area

The positive prompt is split into four parts for convenience. All parts are merged before generation unless Anima Mod Guidance changes the handling of the meta-tag section.

• positive_1: artist tags and LoRA trigger words. This is where the selected style preset text is usually inserted.
• positive_2: meta / quality tags. This is also the input used by Anima Mod Guidance, so keep it separate from the normal prompt.
• positive_3: the main prompt. In normal use, this is the part you edit most often.
• positive_4: trailing quality or composition text.
• negative / negative_extra: the negative prompt is also split so extra exclusions can be added without rewriting the main negative prompt.

If you do not like the split prompt structure, you can bypass it and send a single full prompt directly into Set_promp_t_pos.

1. LoRA Stack and Style Presets

The workflow uses a two-stage LoRA stack.

• LoRA_before: LoRAs that should always be applied.
• style profiles: the profile select / style_num value selects one of the prepared style profiles.
• Style_TEXT: the selected style profile also adds its matching artist tags and trigger words into the prompt.
• LoRA_after: the selected preset LoRA stack is passed forward and later converted into a metadata-friendly string for Image Saver.

This means you can keep common LoRAs in the always-on stack and switch between style profiles by changing only the profile number.

2. Model Load: Set ANIMA Models

The Set ANIMA Models group loads the Anima diffusion model, Qwen text encoder, VAE, LoRA stack, prompt conditioning, sampler settings, Sage Attention, Anima Mod Guidance, and Torch Compile settings into a shared ctx_ANIMA context.

• ANIMA load: choose the Anima diffusion model.
• clip_name: use the Qwen text encoder required by the Anima workflow.
• vae_name: use the Qwen image VAE used by the workflow.
• shift: the workflow uses the Anima default shift value of 3.
• Sage_Attention: set to auto by default.
• Use TorchCompile: enable this for speed after the first compile pass. Disable it first when debugging model patch errors.
• disable_dynamic_vram: keep this enabled when using compile to avoid dynamic VRAM related instability.

3. Anima Generation: Ksampler +

The first sampler is the main generation stage. The workflow uses er_sde with sgm_uniform in the main model settings, and the sampler group can apply DCW and Spectrum before sampling.

• denoise(only i2i): this matters mainly when i2i is enabled. For pure t2i, leaving it as-is is fine.
• Use DCW Node: applies the DCW model patch before sampling.
• Use Spectrum Node: applies Spectrum acceleration before sampling.
• Use i2i: this is passed from the i2i group so the sampler knows whether it is working from an image latent.

If the sampler fails after a model patch, disable Torch Compile and Spectrum first. HighRez can also cause size-related issues when the image size is not compatible.

4. HighRez

The HighRez group performs a second-pass upscale/refinement. In v5.1 it includes an option to convert the latent into an image before upscaling. Keeping this option enabled is recommended because very low denoise values can otherwise make the image look mushy.

• Use HighRez: turns the second pass on or off.
• HighRez_scale_by: controls the scale multiplier. The workflow commonly uses a modest multiplier such as 1.25.
• Denoise Strength: controls how much the second pass changes the image.
• latent to image conversion: recommended to keep enabled.

Be careful with image sizes. HighRez works best when the resulting resolution stays compatible with ComfyUI's expected size multiples. If you use i2i, HighRez is often not recommended because resolution scaling can become harder to control.

5. Inpainting: Anima LLLite

The inpainting group uses Anima LLLite. It is designed to work by drawing a mask in the PreviewBridge area and rerunning only the inpaint path.

• Use LLLite: turns inpainting on or off.
• PreviewBridge: use this area to draw or edit the mask.
• Using New Seed: enable this if you want a separate seed for the inpaint result.
• positive / negative: use the inpainting prompt area to describe only what should be changed or repaired.
• Use Anima Mod Guidance: if enabled, the workflow handles the meta-tag section automatically, so you do not need to manually repeat positive_2 in the inpainting prompt.

Place the Anima LLLite inpainting model in ComfyUI/models/controlnet/.

6. SAM3 Detailers

The workflow includes separate detailer groups for face and eyes. They use SAM3 to detect the target part, convert it into SEGS, and then run a localized detail pass.

• Use Detailer: turns each detailer block on or off.
• Detect part: examples are face and eyes.
• Detect num: controls how many detected parts are processed. It is connected to the character count helper.
• SAM3_threshold / refine_iterations: adjust these if masks are too loose or too strict.
• SEGS_crop_factor: controls how much surrounding area is included around the detected part.
• Detailer_denoise: lower values preserve the original face/eyes; higher values redraw more aggressively.
• Use DCW Node / Use Spectrum Node: optional patches for the detailer pass.

The original guide treats this as an optional feature. If you do not need face or eye repair, leave it off.

7. USDU Upscale

The USDU Upscale group is intended for larger images, especially when the final size is above roughly 2048 pixels on a side or when tile-based refinement is safer than a normal HighRez pass.

• Using USDU: turns USDU upscaling on or off.
• Upscale_scale: controls the final scale.
• tiles num: controls how many tiles are used per side.
• denoise: keep this low if you only want enhancement rather than a redraw.
• Use DCW Node / Use Spectrum: optional acceleration or guidance patches for the upscale pass.

8. Image Saver and Civitai Metadata

The workflow uses Image Saver to store generation metadata in a Civitai-friendly way. The LoRA stack is converted into a prompt-style LoRA string for metadata, and additional resource hashes can be added through the hash input.

• Keep metadata saving enabled if you want Civitai to recognize the workflow resources.
• Do not delete prompt metadata before upload. If you do not want it visible, hide it on Civitai instead.
• Use the additional hash input when you need to manually add a Civitai resource that is not automatically detected.
• Filename / path: change the version folder text, such as V19, if you want to keep outputs separated by workflow version.

Anima Mod Guidance

Anima Mod Guidance is a ComfyUI implementation of modulation guidance for Anima models. In simple terms, it adds an extra CLIP-based guidance signal to the Anima diffusion path. It uses the difference between positive and negative CLIP pooled embeddings and injects that signal into the model.

In this workflow, the important practical rule is simple:

• positive_2 is for meta tags. Do not mix it with the main subject prompt.
• When Use Anima Mod Guidance is enabled, the workflow sends positive_2 to the Mod Guidance node and removes it from the normal merged prompt to avoid double-applying the same quality tags.
• Use quality_tags and mod_w_profile in the model and inpaint groups to control how strongly this guidance is applied.

Expected effects can include better tag following, better character adherence, reduced color leakage, and sometimes improved visual quality. It may also weaken style strength in some cases, so compare outputs with the toggle on and off when building a new preset.

NAIA Prompt Assistance

The NAIA group can overwrite positive_3 with the prompt generated by NAIA. Earlier versions could break workflow caching when NAIA was used, but this version is arranged so that inpainting can still be reused more comfortably.

• Turn on the top NAIA toggle to use it.
• NAIA output replaces the normal positive_3 section.
• If the node is missing after Manager installation, install comfyui-naia-bridge manually.

i2i Usage

1. Load an image in the i2i image input.
2. Turn on Use i2i.
3. Set an appropriate image size or megapixel value based on the input image.
4. Make sure the VAE is set correctly.
5. Adjust the first sampler denoise value. Higher denoise changes the source image more strongly.
6. For i2i, avoid HighRez unless you are carefully controlling the final resolution.

Recommended Troubleshooting Order

1. Check missing custom nodes and update the custom nodes first.
2. Turn off Spectrum in the sampler / HighRez / detailer / USDU blocks.
3. Turn off Torch Compile.
4. Turn off HighRez and run the base sampler only.
5. Turn off detailers and USDU.
6. Check whether the image size is compatible with the required size multiples.
7. If i2i or inpainting fails, check that the VAE and mask are correctly connected.

Useful Links

• ComfyUI-Spectrum-KSampler
• ComfyUI-Spectrum-sdxl
• ComfyUI-DCW
• NAIA Bridge
• ComfyUI-Anima-LLLite
• Anima Mod Guidance for ComfyUI
• SageAttention

Notes

• This workflow does not require every optional block to be enabled. For basic generation, the model load, prompt, LoRA stack, first sampler, and image saver are enough.
• HighRez, detailers, inpainting, and USDU are optional post-processing paths.
• The prepared style profiles are only examples. You can replace them with your own LoRA combinations and artist tags.
• When sharing images on Civitai, keep the metadata generated by Image Saver so Easy Remix and resource detection can work more reliably.