<h2>🚀 Drag-and-Drop ComfyUI Workflow Included</h2>
Every version now comes with <code>AnimaEdit_Reference_Instruction_Workflow.json</code> attached directly in the Files section! Simply download the workflow and drop it into ComfyUI.
<ul>
 <li>Primary Download (Default): <code>anima_edit_v*_anima_ref_adapter_native.safetensors</code> — Pre-converted with <code>diffusion_model.blocks</code> keys and embedded <code>arch: asym</code> metadata for instant plug-and-play in stock <code>ComfyUI-Anima-Ref-Adapter</code>.</li>
 <li>Secondary File: <code>anima_edit_v*_standard_kohya_lora.safetensors</code> — Retains raw Kohya training keys for custom scripts and standard LoRA loaders.</li>
 <li>CRITICAL SAMPLING RULES: On Anima Turbo models, set CFG: 1.0 (Strict) and Steps: 6 with Euler sampler. Never use CFG &gt; 1.0 on distilled Turbo checkpoints or the latent math will diverge and produce a blank white image!</li>
</ul>

<h2>⚠️ Forge Neo Support Status &amp; Architecture Breakdown</h2>
Support for this specific model suite in Forge Neo is unverified at this time:
<ul>
 <li>Cosmos Ref Latent (v1.0): <a target="_blank" rel="ugc" href="https://github.com/Haoming02/sd-webui-forge-classic/tree/neo">Forge Neo</a> currently supports the original Anima Edit setup, which uses Cosmos Reference Latent (temporal frame concatenation of the reference image: <code>torch.cat([z_t, z_ref], dim=2)</code>). Because our v1.0 checkpoint was trained using this exact temporal frame-concat approach, v1.0 may be compatible with Forge Neo.</li>
 <li>Asymmetric KV-Cache (v2.0 – v3.0): Our higher-tier iterations (v2.0 through v3.0) use a fundamentally different, decoupled Asymmetric Cross-Attention KV-Caching architecture. This allows extreme identity preservation, independent RoPE positioning, and supports different aspect ratios and resolutions between the reference and output images. Forge Neo does not natively have custom hooks for this asymmetric KV-conditioning yet.</li>
 <li>Roadmap: We are actively investigating ways to support v2.0–v3.0 within Forge Neo, and may develop a dedicated Forge Neo extension for asymmetric KV support soon. Until then, ComfyUI with the official <code>ComfyUI-Anima-Ref-Adapter</code> is the verified and recommended environment.</li>
</ul>

<hr />

<h1>AnimaEdit: Reference &amp; Instruction Image Editing LoRA</h1>


  AnyAnima (Anima 2B) 
  85,000 Steps · Batch 4 
  90k Paired Dataset 
  Multi-Checkpoint Progression 
  Turbo / Distilled Ready 


AnimaEdit is a high-speed, instruction-guided image-to-image and reference-driven editing suite trained on AnyAnima (Anima 2B DiT). It executes pinpoint localized modifications—clothing swaps, expression shifts, hairstyle edits, pose transformations, and background replacements—while locking character facial identity and stylistic fidelity.

Extensively tested and compatible with standard diffusion baselines as well as fast low-step turbo / distilled models. All exact sampling setups, guidance scales, and step configurations are directly inspectable in the generation metadata embedded within the showcase preview cards.

<hr />

<h2>🧩 ComfyUI Custom Node Integration</h2>

To run these checkpoints in ComfyUI, install the official adapter repository:


 👉 <a target="_blank" rel="ugc" href="https://github.com/ElectronGirl/ComfyUI-Anima-Ref-Adapter">GitHub: ElectronGirl/ComfyUI-Anima-Ref-Adapter</a>


<ul>
 <li>
 KV-Cache Cross-Attention Models (v2.0 – v3.0):
 Use the <code>AnimaRefAdapter</code> node for decoupled reference conditioning. Does not inflate sequence length.
 </li>
 <li>
 Bidirectional Temporal Baseline (v1.0):
 Use the <code>AnimaRefLatent</code> node for frame-concatenation workflows.
 </li>
</ul>

<hr />

<h2>📦 Model Progression &amp; Checkpoint Guide</h2>

The release is split into trait-focused iterations so you can pick the exact balance of instruction strength vs reference locking:

<ul>
 <li>
 v3.0 — Strong (Distractor-Suppressed)
  RECOMMENDED / PEAK 
 
 Final checkpoint trained through 85,000 steps. Features active background distractor suppression and initial tuning for dual reference inputs. Delivers the strongest prompt adherence and cleanest separation between foreground subject and scene edits.
 </li>
 
 <li>
 v2.5 — Precise (Multi-Resolution)
 
 Trained to 75,000 steps with multi-resolution scaling. Excels at intricate localized changes like clothing seams, accessories, and jewelry without bleeding into the background.
 </li>
 
 <li>
 v2.2 — Focused (Tight Identity Lock)
 
 Trained to 65,000 steps. Emphasizes tight preservation of original facial features and character drawing style. Ideal for subtle expression adjustments.
 </li>
 
 <li>
 v2.1 — Dynamic Balance
 
 Trained to 57,000 steps. Provides compositional freedom for pose alterations, camera angle changes, and character movement.
 </li>
 
 <li>
 v2.0 — Natural Balance
 
 Trained to 50,000 steps. Foundational milestone for KV-cache attention stream conditioning. Produces soft, natural lighting and atmospheric color blending.
 </li>
 
 <li>
 v1.0 — Temporal Frame-Concat (Non-KV)
 
 Trained to 17,000 steps using temporal frame concatenation. Provided as an architectural reference for frame-concat ComfyUI workflows.
 </li>
</ul>

<hr />

<h2>🎯 Prompting Strategies (3 Supported Modalities)</h2>

The training pipeline exposed the model to three distinct prompt structures—use whichever fits your workflow:

<ol>
 <li>
 Imperative Instruction:
 Direct commands specifying the exact modification.
 
 <code>"change clothing to a navy blue school uniform blazer, keep hair and expression same"</code>
 </li>
 
 <li>
 Conversational / Descriptive:
 Contextual natural language explaining the intended delta.
 
 <code>"make the character smile brightly with eyes closed, keeping the beach background untouched"</code>
 </li>
 
 <li>
 Minimal Keyword Triggers:
 Compact tag sequences for automated batch pipelines.
 
 <code>"winking, happy, blush, same character, same pose, same clothing"</code>
 </li>
</ol>

<hr />

<h2>🎨 Supported Edit Task Categories</h2>

Trained on 90,000 verified image pairs (80,000 synthetic AnyAnima pairs with multi-pass filtering + 10,000 external edit pairs) across eight specialized tasks:

<ul>
 <li>Background Replacement: Swap exterior environments, weather, time of day, and scenic architecture.</li>
 <li>General Scenery &amp; Environmental Edit: Modify localized lighting, room interiors, foliage, and ambient props.</li>
 <li>Pose Alteration: Reposition limbs, hands, head turns, and dynamic postures.</li>
 <li>Clothing &amp; Costume Swap: Complete outfit replacement, textile changes, and uniform swaps.</li>
 <li>Hairstyle Manipulation: Adjust hair length, twin-tails, buns, braids, and hair color.</li>
 <li>Expression Adjustment: Smile, wink, surprise, blush, focused gaze, and emotional shifts.</li>
 <li>Character Insertion: Introduce consistent anime subjects into established backgrounds.</li>
 <li>Background Extraction &amp; Subject Isolation: Cleanly isolate character cutouts or extract clean scenery plates.</li>
</ul>

<hr />

<h2>⚠️ Coordinate Anchoring &amp; Technical Quirks</h2>

<h3>1. Aspect Ratio Shifts on In-Place Edits (Top-Left Anchoring)</h3>
The model was trained natively at 512x512 resolution with variable-resolution input conditioning, and easily scales to higher resolutions (768x768, 1024x1024, etc.).
However, when performing an in-place edit while simultaneously changing the output aspect ratio (e.g., expanding from 512x512 square to 576x384 or 768x512 wide), the DiT 2D RoPE position embedding anchors coordinates to the top-left origin (0,0).
This causes a centered subject to stay anchored toward the left side of the widened canvas, with new background scenery generated on the right side. If you want centered framing when expanding canvas dimensions, pad or re-center your reference image prior to generation.

<h3>2. Dual-Reference Conditioning (v3.0)</h3>
v3.0 includes initial tuning for dual reference inputs with distractor suppression to isolate the primary subject. While effective at discarding visual noise, multi-reference composition remains experimental.

<hr />

All progression checkpoints, metadata, and preview verification cards are published under open creative use. Happy generating!

AnimaEdit: Reference & Instruction Image Editing LoRA

v1.0 (Temporal Frame-Concat): Trained for 17,000 steps using temporal bidirectional frame concatenation (non-KV stream).Specifically designed for workflows utilizing the <code>AnimaRefLatent</code> node rather than decoupled asymmetric attention caching.

Files with SHA256: 91326be636494afc9360704d6bc7e7c0bba01c19038a8f58a6e80db1e5172573