<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e968b7d3-3324-43e7-ad13-acbaa45a188a/original=true/e968b7d3-3324-43e7-ad13-acbaa45a188a.jpeg" />ATTACH V2A new function in the node pack. Built to raise the influence of Qwen 3.5 4B. It needs a little more memory than the old method. Direct development of <a target="_blank" rel="ugc" href="https://civarchive.com/models/2964040/delta-qwen-35-4b">Delta Qwen 3.5 4B</a>. Source: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2455272/anima-2b-qwen-35-4b-text-encoder?modelVersionId=2760745">Anima 2B — Qwen 3.5 4B Text Encoder</a>.This model has two modes. Both are selected in the new Clip loader.3.8 1.1 base — the old mode, unchanged.Qwen 3.5 strong — new. Slightly better frame reading, better scene layout. Test it.Important: strong is still a test mode. It has its own problems, and a number of them are already known. If you hit one yourself, or have a fix in mind, write it in the comments.Delta Mix 3.8B Qwen 3.5 - v5 Same line. v5 is the Attach V2 runtime. The bake is still DiT only. The Semantic Connector is still a sidecar, not inside the file.The working path is the earlier one: 3.8 1.1 base, full frame. Same sidecar, same official connector path, same picture. If you only want what the previous version already did, stop here.qwen 3.5 strong is the other CLIP format. It makes Qwen 3.5 4B the leading CLIP. Native 0.6B stays in the graph. This format does not use the 2560 sidecar as the conductor.Samplers: dpmpp_2m / euler_a / res_multistep Schedulers: res_uniform / normal / betaWarningAttach V2 is not stable when you switch CLIP format on one live load. The problem is known. Part of it cannot be fixed in this node without breaking the path.If the frame smears, drifts, or ignores the format you just set: change the loaded model, then load this checkpoint again, then set the format you want. Do not keep switching base / strong on the same load.Delta Mix 3.8B Qwen 3.5 - v4A recut of the same line, not a new family. v3 tightens frame stability and edge clarity: lines and object borders stay even instead of swimming. Same dual brief — human and furry.Samplers: dpmpp_2m / euler_a / res_multistep Schedulers: res_uniform / normal / betaDelta Mix 3.8B Qwen 3.5 - v3 Anatomy and hands: fewer collapsed fingers and joint errors on the same test prompts. Q35 path: the base holds pose / crop more steadily when the 4B encoder is in the graph. Stem (W): replaced. W is now a custom assembled 2B stack, not the previous floor.Delta Mix 3.8B Qwen 3.5 - v2Anima 3.8B mix for the official v1.1 chassis and Qwen 3.5 Connector v2.Source: CircleStone Labs / lylogummy Anima 3.8B v1.1.This bake fills W + Gen-1 + Gen-2 (Delta Mix overlay).Connector v2 and the Q35 watch slots stay as in v1.1.Qwen 3.5 is attached from a sidecar, not baked into this file.Cleaner image than v1. Fewer W / Gen-2 clashes.Fine detail and extra fingers still slip — for v3.Delta Mix 3.8B Qwen 3.5 -v1Anima 3.8B mix for the official v1.1 chassis and Qwen 3.5 Connector v2. Source: CircleStone Labs / lylogummy Anima 3.8B v1.1.The DiT is a Delta Mix overlay (stem + light Gen-1). Gen-2 insert slots and the Connector stay as in v1.1. Qwen 3.5 is attached from a sidecar, not baked into this file. <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f90fd88a-a139-4dcd-a6b1-dd25d03dbdd5/original=true/f90fd88a-a139-4dcd-a6b1-dd25d03dbdd5.jpeg" /> Downloads<ol><li>This checkpoint</li><li>Sidecar anima_connector_only → same folder (or !Anima3.8/)</li><li>Nodes: <a target="_blank" rel="ugc" href="https://github.com/bestluner-create/ComfyUI-Anima-Delta-Q35v2.git">ComfyUI-Anima-Delta-Q35-V2</a></li><li>Official runtime: <a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B (MIT</a>) — Qwen 3.5 4B loader and Connector class</li><li>Text encoders from the v1.1 guide: Qwen 0.6B + Qwen 3.5 4B</li><li><a rel="ugc" href="https://github.com/GumGum10/comfyui-qwen35-anima.git">Qwen 3.5 4B loader:</a> GumGum10/comfyui-qwen35-anima. CLIP Loader V2 calls it for qwen 3.5 strong. Weights page: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2455272">https://civarchive.com/models/2455272</a></li><li>Lora loader: <a target="_blank" rel="ugc" href="https://github.com/bestluner-create/ComfyUI-Anima-Delta-Q35-LoRA.git">ComfyUI-Anima-Delta-Q35-LoRA</a></li></ol><pre><code>UNET → DeltaMix Q35 LoRA → Attach V2 → Prompt V2</code></pre><h2 id="install">Install</h2>This bake is DiT only (~7 GB). The Connector and Qwen 3.5 4B are not inside the file.<ol><li>This checkpoint → ComfyUI/models/diffusion_models/ (or models/unet/)</li><li>anima_connector_only.safetensors → same folder</li><li>qwen35_4b.safetensors → ComfyUI/models/text_encoders/ (or models/clip/)</li><li>qwen_3_06b_base.safetensors → same folder. CLIPLoader, type stable_diffusion</li><li>ComfyUI-Anima-Delta-Q35-V2 → ComfyUI/custom_nodes/</li><li>Both GumGum10 runtimes, same folder:</li></ol>Also needed: native CLIP qwen_3_06b_base, same place as the 0.6B line.<h3 id="2.-nodes-(git)">2. Nodes (Git)</h3><pre><code>ComfyUI/custom_nodes/comfyui-anima-3-8B
 https://github.com/GumGum10/comfyui-anima-3-8B

ComfyUI/custom_nodes/ComfyUI-Anima-Delta-Q35-V2
 (zip on this page) or git clone https://github.com/bestluner-create/ComfyUI-Anima-Delta-Q35v2.git

ComfyUI/custom_nodes/ComfyUI-Anima-Delta-Q35-LoRA
 https://github.com/bestluner-create/ComfyUI-Anima-Delta-Q35-LoRA</code></pre>Restart ComfyUI after the three packs.comfyui-anima-3-8B is the official runtime (4B loader + Connector class). Without it the Q35 graph will miss nodes.Graph<pre><code>UNETLoader this bake (~7.3G)
 → DeltaMix Q35 Attach V2 sidecar on 3.8 1.1 base
CLIP Loader V2 clip_format
 → DeltaMix Q35 Prompt V2 CLIP 0.6B + CLIP 4B
 expanded → CFG / Sampler</code></pre>LoRA (optional):<pre><code>UNET → DeltaMix Q35 LoRA → Attach V2 → Prompt V2</code></pre>Prompt Dual Delta: pos = quality tags + scene, neg = negative tags only. Use expanded for sampling. native is 0.6B-only.Extract once: DeltaMix Q35 Extract Connector on anima38B_11.safetensors. Do not extract again unless v1.1 changes.This file does not include the Connector (~1.2G) or the 4B encoder. Do not load *Q35bundle in this graph. Bundles are for the official v1.1 loader. SamplerSame stack as the v1.1 graph: RandomNoise → CFGGuider → SamplerCustomAdvanced + BasicScheduler. A plain KSampler is fine if the numbers match.Steps 40. CFG 6. Denoise 1.0. Sampler: er_sde, Euler, or whatever you already run on 3.8. Scheduler: normal for this mix. beta if you want the stock v1.1 cadence.Empty latent 1024×1024. 1920×1080 is stable.Connector strength 1.0. Both Prompt nodes use the expanded output.Do not match seeds between the official V2Loader graph and this one. The noise path is different. Compare seeds only inside the same graph. <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/54feb12e-be8d-418c-80df-24bddf939ccf/original=true/54feb12e-be8d-418c-80df-24bddf939ccf.jpeg" /> CreditsV-5<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima">Anima</a> - CircleStone Labs</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2880480/anima-38b">3.8B v1.1</a> chassis and Semantic Connector v2 - lylogummy</li><li><a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B</a> - GumGum10 (MIT)</li><li>Extract / Attach / Prompt - Delta Mix</li><li>Method: Anima <a target="_blank" rel="ugc" href="https://civarchive.com/models/2908374/anima-delta-mix-2b-into-29b-38b">Delta Mix — 2B into 2.9B / 3.8B</a></li><li>Custom - G1</li><li>Custom - G2</li><li>Custom - core</li></ul>V-4<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima">Anima</a> - CircleStone Labs</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2880480/anima-38b">3.8B v1.1</a> chassis and Semantic Connector v2 - lylogummy</li><li><a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B</a> - GumGum10 (MIT)</li><li>Extract / Attach / Prompt - Delta Mix</li><li>Method: Anima <a target="_blank" rel="ugc" href="https://civarchive.com/models/2908374/anima-delta-mix-2b-into-29b-38b">Delta Mix — 2B into 2.9B / 3.8B</a></li><li>Custom - G1</li><li>Custom - G2</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2945343/demix">DeMix</a> - core</li></ul>V-3<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima">Anima</a> - CircleStone Labs</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2880480/anima-38b">3.8B v1.1</a> chassis and Semantic Connector v2 - lylogummy</li><li><a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B</a> - GumGum10 (MIT)</li><li>Extract / Attach / Prompt - Delta Mix</li><li>Method: Anima <a target="_blank" rel="ugc" href="https://civarchive.com/models/2908374/anima-delta-mix-2b-into-29b-38b">Delta Mix — 2B into 2.9B / 3.8B</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2795320/aurarising">Aura_Rising</a> - G1</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2726336/animurr?modelVersionId=3153097">aniMurr</a> - G2</li><li>Custom - core</li></ul> V-2<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima">Anima</a> - CircleStone Labs</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2880480/anima-38b">3.8B v1.1</a> chassis and Semantic Connector v2 - lylogummy</li><li><a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B</a> - GumGum10 (MIT)</li><li>Extract / Attach / Prompt - Delta Mix</li><li>Method: Anima <a target="_blank" rel="ugc" href="https://civarchive.com/models/2908374/anima-delta-mix-2b-into-29b-38b">Delta Mix — 2B into 2.9B / 3.8B</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2795320/aurarising">Aura_Rising</a> - G1</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2726336/animurr?modelVersionId=3153097">aniMurr</a> - G2</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima?modelVersionId=3126581">anima_aesthetic</a> - core</li></ul> V-1<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima">Anima</a> - CircleStone Labs</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2880480/anima-38b">3.8B v1.1</a> chassis and Semantic Connector v2 - lylogummy</li><li><a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8B</a> - GumGum10 (MIT)</li><li>Extract / Attach / Prompt - Delta Mix</li><li>Method: Anima <a target="_blank" rel="ugc" href="https://civarchive.com/models/2908374/anima-delta-mix-2b-into-29b-38b">Delta Mix — 2B into 2.9B / 3.8B</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2795320/aurarising">Aura_Rising</a> - G1</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2839674/chosen-mixanima">Chosen-mix_anima</a> - core</li></ul>

Anima Delta Mix 3.8B Qwen 3.5

animaDeltaMix38BQwen3_v10.json

animaDeltaMix38BQwen3_v10_3180544.safetensors

animaDeltaMix38BQwen3_v10_txt.safetensors

animaDeltaMix38BQwen3_v10_3180553.safetensors

v1.0

DeltaMix3.8B.1.1,V5

DeltaMixV4Q35

DeltaMix3.8B_1.1_V3

DeltaMix3.8B_1.1_V2

What changed from V2 • Anatomy and hands: fewer collapsed fingers and joint errors on the same test prompts. • Q35 path: the base holds pose / crop more steadily when the 4B encoder is in the graph. • Stem (W): replaced. W is now a custom assembled 2B stack, not the previous floor.

animaDeltaMix38BQwen3_deltamix38B11V3.zip

animaDeltaMix38BQwen3_deltamix38B11V3.json

animaDeltaMix38BQwen3_deltamix38B11V3_3210928.safetensors

animaDeltaMix38BQwen3_deltamix38B11V3_fp8.safetensors

animaDeltaMix38BQwen3_deltamix38B11V3_3205844.safetensors

Anima 3.8B

anima38B_base_3138693.safetensors

anima38B_base_txt.safetensors

anima38B_base_3138681.safetensors

BASE

animaDeltaMix38BQwen3_deltamix38B11V5_3270795.safetensors

animaDeltaMix38BQwen3_deltamix38B11V5.json

animaDeltaMix38BQwen3_deltamix38B11V5.zip

animaDeltaMix38BQwen3_deltamix38B11V5_3270789.safetensors

animaDeltaMix38BQwen3_deltamix38B11V5_fp8.safetensors

animaDeltaMix38BQwen3_deltamixv4q35_3224706.safetensors

animaDeltaMix38BQwen3_deltamixv4q35_3224701.safetensors

animaDeltaMix38BQwen3_deltamixv4q35.json

animaDeltaMix38BQwen3_deltamixv4q35_trainingData.zip

animaDeltaMix38BQwen3_deltamixv4q35_fp8.safetensors

<h3 id="delta-mix-qwen-3.5-4b">Delta Mix Qwen 3.5 4B</h3>Dual-encoder routing for Anima 2B / 2.9 / 3.8-0.6B.Keeps the native Qwen 0.6B clip and adds Qwen 3.5 4B only where you ask it to.This is glue, not a new text encoder. The 4B loader and weights stay with the original pack.A separate page for the method is available at <a target="_blank" rel="ugc" href="https://civarchive.com/models/2964040/delta-qwen-35-4b?modelVersionId=3358244">https://civarchive.com/models/2964040/delta-qwen-35-4b?modelVersionId=3358244</a><h3 id="what-it-does">What it does</h3>Two encodes of the same prompt: stock 0.6B on the edges, 4B in the middle.Default mode <code>blocks_ortho</code> projects 4B toward native space, clips per-token spikes <code>delta_guard</code>), then injects a soft block window.Start at strength 0.50 / delta_cap 0.50 / guard on. That is a visible start, not an optimum. The checkpoint moves more than the sliders.<h3 id="works-on-2b-(28)-2.9-(40)-and-3.8-0.6b-(52).-not-the-3.8-q35-connector-pack.">Works on 2B (28), 2.9 (40), and 3.8-0.6B (52). Not the 3.8 Q35 connector pack.</h3><h3 id="downloads">Downloads</h3>This pack (nodes + example workflow)- GitHub: <a target="_blank" rel="ugc" href="https://github.com/bestluner-create/Delta-Mix-Qwen-3-5-4B">https://github.com/bestluner-create/Delta-Mix-Qwen-3-5-4B</a>- Same zip and workflow are attached on this CivitAI page under Files.Required — 4B loader (not bundled)- Nodes: <a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-qwen35-anima">https://github.com/GumGum10/comfyui-qwen35-anima</a>- Idea / original 4B-on-Anima page: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2455272/anima-2b-qwen-35-4b-text-encoder">https://civarchive.com/models/2455272/anima-2b-qwen-35-4b-text-encoder</a>- Weights + sidecar: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b">https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b</a>Put <code>qwen35_4b.safetensors</code> in <code>models/text_encoders/</code>.Put <code>rotation_matrix.safetensors</code> and <code>qwen35_tokenizer/</code> in <code>custom_nodes/comfyui-qwen35-anima/</code>.Required — native Anima TE- <code>qwen_3_06b_base.safetensors</code> → <code>models/text_encoders/</code>(stock Anima CLIP, CLIPLoader type <code>qwen_image</code> or <code>stable_diffusion</code>)Install both custom_node folders, restart Comfy, load the workflow from Files.Required for 2.9 / 40-block checkpointsWithout this patch ComfyUI still loads the file as 28 blocks and silently drops 12 layers.<a target="_blank" rel="ugc" href="https://github.com/gazingstars123/ComfyUI-Anima-2.9B">https://github.com/gazingstars123/ComfyUI-Anima-2.9B</a>Same text → CLIPTextEncode 0.6B (edge) and CLIPTextEncode 4B (center), pos and neg.LoRA on MODEL only. Do not LoRA the 4B CLIP. Do not lerp the two conds before the node.LoadQwen35AnimaCLIP: alignment ON, 0.5. <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/af62c90d-dc38-48d9-8de7-08a0c1c7b107/original=true/af62c90d-dc38-48d9-8de7-08a0c1c7b107.jpeg" /><h3 id="(cap-strip)">(cap strip)</h3>Same seed, same prompt, strength locked at 1.00.1 Base — node off, native 0.6B only.2 cap 0.25 — 4B on, tight per-token clip.3 cap 0.50 — default cap, strength still 1.00.4 cap 0.75 — clip loosened.delta_cap is not “more style”.It is the per-token ceiling on |C_4B − C_0.6B|.Low cap keeps the lock (goggles, heart, stance).High cap at strength 1.00 lets 4B rewrite pose and outfit.Raise cap only after strength is already set.delta_cap — how large a single token residual may be versus native 0.6B.Default 0.50. Lower = closer to Base. Higher = 4B may move props, pose, clothes.Does nothing if delta_guard is off. Pair with strength; do not slam both to 1.00.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7bcf9d28-cc8a-477a-99c0-19b5e1a2407d/original=true/7bcf9d28-cc8a-477a-99c0-19b5e1a2407d.jpeg" /><h3 id="(strength-strip)">(strength strip)</h3>Same seed, same prompt, delta_cap locked at 0.50.1 s 0.25 — 4B barely in.2 s 0.50 — pack default.3 s 0.75 — pose starts to sit.4 s 1.00 — sleeves and stance rewrite.strength is how much of the already-clipped 4B residual is added.cap decides the ceiling; strength decides how far you walk toward it.0.25–0.50 keeps the lock. Past 0.75 the camera and clothes move first.strength — mix amount after guard: C = C_0.6B + strength × D_guarded.Default 0.50. This is the main “how much 4B” knob.Leave cap at 0.50 while you find a strength that still holds the pose.<pre><code>CheckpointLoader Anima 2B / 2.9 / 3.8-0.6B
CLIPLoader qwen_3_06b_base.safetensors
 type: qwen_image (or stable_diffusion)
LoadQwen35AnimaCLIP qwen35_4b.safetensors
 alignment ON, strength 0.5
 calibration OFF, output_scale 1.0
VAELoader qwen_image_vae.safetensors
EmptySD3LatentImage your size

same POSITIVE text ──► CLIPTextEncode (0.6B) ──► cond_edge_pos
 └─► CLIPTextEncode (4B) ──► cond_center_pos

same NEGATIVE text ──► CLIPTextEncode (0.6B) ──► cond_edge_neg
 └─► CLIPTextEncode (4B) ──► cond_center_neg

LoraLoader MODEL only
 CLIP slot = 0.6B if the LoRA has TE keys
 do not attach LoRA to the 4B CLIP

DeltaQ4B Apply
 model ← Checkpoint / LoraLoader
 cond_edge_pos ← 0.6B positive
 cond_center_pos ← 4B positive
 cond_edge_neg ← 0.6B negative
 cond_center_neg ← 4B negative
 mode blocks_ortho
 strength 0.50
 delta_guard True
 delta_cap 0.50
 center_start/end 8/19 (2B)
 10/29 (2.9)
 14/37 (3.8)
 taper 4
 enabled True

KSampler ← model / positive / negative from DeltaQ4B Apply
VAEDecode ← KSampler + VAE
SaveImage</code></pre><h3 id="2b-28-blocks-8-19">2B 28 blocks → 8–19</h3><h3 id="2.9-40-blocks-10-29">2.9 40 blocks → 10–29</h3><h3 id="3.8-52-blocks-14-37">3.8 52 blocks → 14–37</h3><h3 id="inset-1-2-from-the-edges-taper-4">inset 1–2 from the edges, taper 4</h3><h3 id="compatibility">Compatibility</h3>The node runs on Anima 2B (28 blocks), 2.9B (40) and 3.8B-0.6B (52).Change the center window to match the stem. It is not the 3.8 Q35 connector pack.Sample cards on this page use a 2.9-based merge.<h3 id="credits">Credits</h3>Anima 2.9B — architecture and official 2.9 checkpoint[gazingstars321 / Anima-2.9B](<a target="_blank" rel="ugc" href="https://civarchive.com/models/2855007/anima-29b">https://civarchive.com/models/2855007/anima-29b</a>)4B-on-Anima idea — original method this pack routes around[LyloGummy / Anima 2B Qwen 3.5 4B Text Encoder](<a target="_blank" rel="ugc" href="https://civarchive.com/models/2455272/anima-2b-qwen-35-4b-text-encoder">https://civarchive.com/models/2455272/anima-2b-qwen-35-4b-text-encoder</a>)Anima itself is CircleStone Labs. 4B loader and weights stay with GumGum10 / lylogummy (linked in Downloads). This page is the dual-encoder glue only.

Delta Mix 2.9B Qwen 3,5 4B

deltaMix29BQwen354B_29Q4B_bf16.safetensors

deltaMix29BQwen354B_29Q4B.zip

deltaMix29BQwen354B_29Q4B.json

deltaMix29BQwen354B_29Q4B_fp8.safetensors

2.9 Q4B

<h1 id="qwen-3.5-4b-text-encoder-for-anima-2b">Qwen 3.5 4B Text Encoder for Anima 2B</h1>NEW → Now supported on Forge Neo (sd-webui-forge-neo) as a native extension! See the Forge Neo install instructions below.<h2 id="installation">Installation</h2><h3 id="comfyui">ComfyUI</h3>Clone the repo into your ComfyUI <code>custom_nodes</code> folder:<pre><code>cd ComfyUI/custom_nodes
git clone https://github.com/GumGum10/comfyui-qwen35-anima.git
</code></pre>Then restart ComfyUI.<h3 id="forge-neo">Forge Neo</h3>Clone or copy the extension into your Forge Neo <code>extensions</code> folder:<pre><code>cd sd-webui-forge-neo/extensions
git clone https://github.com/GumGum10/sd-forge-qwen-35-encoder.git
</code></pre>Then restart Forge Neo. Dependencies (<code>transformers</code>, <code>safetensors</code>) install automatically on first launch.<hr /><h2 id="what-is-this">What Is This?</h2>A drop-in upgrade for Anima 2B's text encoder. The stock Anima ships with a tiny 0.6B parameter text encoder — it works, but it struggles with complex prompts. This replaces it with a 4B parameter encoder that understands your prompts significantly better.The trade-off: the larger encoder needs alignment work to "speak the same language" as the diffusion model. We've done that work and ship the alignment files with this release. You just need to place files in the right folders and toggle a couple of settings.<hr /><h2 id="what-you-get">What You Get</h2>Pros:<ul><li>Much better understanding of complex/long prompts (7× more parameters dedicated to reading your text)</li><li>Better handling of detailed scene descriptions, multiple subjects, and nuanced instructions</li><li>Alignment controls let you blend between raw 4B output and 0.6B-compatible output</li></ul>Cons:<ul><li>Uses more VRAM than the stock 0.6B encoder (~4GB vs ~0.6GB for the text encoder portion)</li><li>Slightly slower encoding (more parameters to run)</li><li>Alignment is an approximation — the diffusion model was trained against the 0.6B, so we're rotating the 4B's output to match. It's very good (0.96 cosine similarity) but not identical</li><li>This is a reverse-engineered implementation — the original author's private code may differ in subtle ways</li></ul><hr /><h2 id="file-placement">File Placement</h2>All files are available at: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b">lylogummy/anima2b-qwen-3.5-4b</a><h3 id="comfyui">ComfyUI</h3>You'll download 4 files:<pre><code>ComfyUI/
├── models/
│ └── text_encoders/
│ └── qwen35_4b.safetensors ← THE TEXT ENCODER WEIGHTS
│
└── custom_nodes/
 └── comfyui-qwen35-anima/ ← THIS CUSTOM NODE FOLDER
 ├── __init__.py ← (comes with the node)
 ├── calibration_params.safetensors ← MAGNITUDE CALIBRATION
 ├── rotation_matrix.safetensors ← ALIGNMENT ROTATION
 └── qwen35_tokenizer/ ← TOKENIZER FILES
 ├── tokenizer.json
 ├── vocab.json
 └── merges.txt
</code></pre><h3 id="forge-neo">Forge Neo</h3>You only need to download 1 file — the calibration files, alignment matrix, and tokenizer are already bundled with the extension:<pre><code>sd-webui-forge-neo/
├── models/
│ └── text_encoder/
│ ├── qwen_3_06b_base.safetensors ← STOCK 0.6B (you already have this)
│ └── qwen35_4b.safetensors ← DOWNLOAD THIS
│
└── extensions/
 └── sd_forge_qwen35_encoder/ ← THIS EXTENSION
 ├── scripts/ ← (comes with extension)
 ├── lib_qwen35/ ← (comes with extension)
 ├── calibration_params.safetensors ← (bundled)
 ├── rotation_matrix.safetensors ← (bundled)
 └── qwen35_tokenizer/ ← (bundled)
</code></pre>Forge Neo note: Keep <code>qwen_3_06b_base.safetensors</code> selected in the top VAE/Text Encoder dropdown — its LLM adapter is still required. Do not put <code>qwen35_4b.safetensors</code> in that top dropdown.<h3 id="where-to-download-each-file:">Where to download each file:</h3>qwen35_4b.safetensors (both ComfyUI and Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/text_encoders">text_encoders/</a> → Place in: <code>ComfyUI/models/text_encoders/</code> or <code>sd-webui-forge-neo/models/text_encoder/</code> → What it does: The actual 4B text encoder model weightscalibration_params.safetensors + rotation_matrix.safetensors (ComfyUI only — bundled in Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/calibration">calibration/</a> → Place in: <code>ComfyUI/custom_nodes/comfyui-qwen35-anima/</code> → What they do: Calibration scales the 4B output to match the 0.6B's magnitude per dimension. The rotation matrix rotates the 4B's concept directions to match what the adapter expects.qwen35_tokenizer/ folder (ComfyUI only — bundled in Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/tokenizer">tokenizer/</a> → Place in: <code>ComfyUI/custom_nodes/comfyui-qwen35-anima/qwen35_tokenizer/</code> → What it does: The correct tokenizer (vocab=248K, NOT the default Qwen3 tokenizer) → Note: This will auto-download from HuggingFace on first use if you don't place it manually.<hr /><h2 id="how-to-use">How to Use</h2><h3 id="comfyui">ComfyUI</h3><ol><li>Add the "Load Qwen3.5 CLIP (Anima)" node (found under <code>loaders → Anima</code>)</li><li>Select <code>qwen35_4b.safetensors</code> from the dropdown</li><li>Connect the CLIP output to a CLIPTextEncode node</li><li>Use with your Anima 2B checkpoint as normal</li></ol><h3 id="forge-neo">Forge Neo</h3><ol><li>Load an Anima 2B checkpoint</li><li>Make sure <code>qwen_3_06b_base.safetensors</code> is in the top VAE/Text Encoder dropdown</li><li>In the generation tab, expand "Qwen3.5 Text Encoder (Anima)" and enable it</li><li>Select <code>qwen35_4b.safetensors</code> in the extension's Model File dropdown</li><li>Generate as normal — the extension intercepts text encoding automatically</li></ol><h3 id="recommended-settings-to-start-(both)">Recommended Settings to Start (both)</h3><pre><code>use_alignment: ON
alignment_strength: 0.5
use_calibration: OFF
output_scale: 1.0
</code></pre>That's it. Generate some images and compare against the stock 0.6B.<hr /><h2 id="tuning-guide">Tuning Guide</h2><h3 id="what-the-settings-actually-do-(plain-english):">What the settings actually do (plain English):</h3>use_alignment — Rotates the 4B's internal "compass" so that when it says "from the side" or "looking up", it points in the same direction the diffusion model expects. Without this, the 4B understands your prompt fine — it just communicates it in a way the diffusion model misreads.alignment_strength (0.0 – 1.0) — The rotation (direction fix) is always on when alignment is enabled. This slider controls how much the magnitude shifts to match the 0.6B:<ul><li>0.0 = Directions fixed, but keep the 4B's own signal strength</li><li>0.5 = Halfway blend ← start here</li><li>1.0 = Fully match the 0.6B's signal strength</li></ul>use_calibration — A finer-grained magnitude adjustment (per dimension instead of uniform). Can help, can also over-correct. Try it on and off and compare.output_scale — A simple multiplier on the final output. Leave at 1.0 unless you know what you're doing.<h3 id="recommended-workflow:">Recommended workflow:</h3><ol><li>Generate with alignment OFF first — see what the raw 4B gives you. The text understanding will be better, but poses/viewpoints may be off.</li><li>Turn alignment ON, set strength to 0.5 — generate the same prompts again. You should see better pose/viewpoint adherence while keeping the 4B's improved understanding.</li><li>Adjust strength — bump it up if spatial stuff is still off, pull it back if quality degrades.</li><li>Optionally enable calibration — compare on/off, keep whichever looks better for your use case.</li></ol><hr /><h2 id="faq">FAQ</h2>Q: Do I need both calibration AND alignment files? A: The alignment file (<code>rotation_matrix.safetensors</code>) is the most important one. Calibration is optional and supplementary. You can use alignment without calibration.Q: Will this work with any Anima 2B checkpoint? A: Yes — any checkpoint built on Anima 2B that uses the standard text encoder pipeline.Q: Does this need extra Python packages? A: For ComfyUI — no, everything ships with ComfyUI already. For Forge Neo — <code>transformers</code> and <code>safetensors</code> install automatically on first launch.Q: How much extra VRAM does this use? A: The 4B encoder weights are FP8 quantized, so roughly ~4GB for the text encoder. The stock 0.6B is under 1GB. Your total VRAM usage depends on your diffusion model + VAE + this.Q: Why not just scale the output by 10× instead of all this alignment stuff? A: Uniform scaling fixes the magnitude but not the directions. The 4B encodes "from the side" as a vector pointing in a completely different direction than the 0.6B. The rotation matrix fixes that. Scaling alone would be like shouting the wrong directions louder.Q: Is this better than the stock 0.6B? A: For text understanding — yes, meaningfully. For raw out-of-the-box image quality — it depends on your alignment settings and prompts. The 0.6B has the advantage of being exactly what the model was trained against. The 4B has the advantage of actually understanding complex prompts. With alignment at 0.5, most users see comparable or better results, especially on detailed prompts where the 0.6B falls short.Q: Can I use this with img2img? A: Yes — works for both txt2img and img2img on both ComfyUI and Forge Neo.Q: Why does Forge Neo still need the 0.6B model loaded? A: The Anima pipeline uses a small LLM adapter that lives on the 0.6B model. This adapter converts text embeddings into the format the diffusion model expects. The 4B provides the text understanding, but the adapter (on the 0.6B) still handles the final conversion. Both models are needed.<hr /><h2 id="credits">Credits</h2><ul><li>Anima 2B: <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima">circlestone-labs</a></li><li>Qwen 3.5 4B for Anima: <a target="_blank" rel="ugc" href="https://huggingface.co/nightknocker/cosmos-qwen3.5">nightknocker/cosmos-qwen3.5</a></li><li>Custom Node, Alignment &amp; Forge Neo Port: <a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-qwen35-anima">GumGum10</a></li></ul>

Anima 2B - Qwen 3.5 4B Text Encoder

anima2BQwen354BText_base.safetensors

Files with SHA256: ea289be7c916726d09953c7db9971c82b280e694b5d7c47f8ad9ffad6acb54ba