Text to Image Models

NSFW Weights for this model are available in Safetensors format. Download them in the Files & versions tab.

Model · Text to image

flux2-klein-GGUF

Chris Cole

9B step-distilled text-to-image — 4 inference steps, ~8s per 1024² image on an A100. One source, three prompts, identical settings — so the difference you see is the prompt. This model's txt2img preset is 4 steps — it is a distilled 4-step model and that is correct for text-to-image. Running img2img at 4 steps produces output indistinguishable from the source at every strength from 0.3 to 0.9: there is not enough schedule left to restyle anything. At 24 steps with guidance 4.0 — the settings this repo's own img2img integration test uses — it restyles properly. The three results above are the same photograph as oil, watercolour and pencil, with the composition intact in each and the medium…

Model · Text to image

Ruach_Studio_Models_v2

Viktor Zhuromskyy

Everything Ruach Studio runs on, in one place and laid out as its own folders: the music model and its decoders, the transcriber, the sliders, the stems' and the sound upscaler's models, Whisper, the style listener, the studio's own adapters, and the Artist's painters with the adapters their authors let be shared. 152 files, 240 GB. A copy of the studio's own disk, file for file: what lies here is what it paints and sings with. It takes the place of goldhub/RuachStudioModels, which goes after 17 October 2026. Nothing here is ours but the conversions and the studio's own adapters. Every other file is its authors' work, unchanged, under its authors' licence, credited below and line by line in…

The MOSS-Image SFT model often draws faint, evenly spaced vertical color stripes across smooth areas like skies, fog, backdrops and walls. This repo has two banding-fix LoRAs that remove them while keeping the image close to what SFT would have made, plus preset workflows that pair the fix with community speed LoRAs for 10–14 step (or 4–8 step) generation. Just want one file that works? MOSS-Image-SFT-DeBand-int8 has the fix (and the Fast preset) merged into ready-to-run 5.4 GB int8 models. No LoRAs or custom nodes. 1. Download the main fix, mossimagesftbandingfixL37-41bf16.safetensors, into ComfyUI/models/loras/. 2. In your MOSS-Image workflow, add LoraLoaderModelOnly after the SFT model…

An Anima fine-tune by silvermoong, using the "Evolved" training strategy with a focus on stable hands, a neutral art style that works with artist tags, and multiple characters. Update ComfyUI or Forge Neo to the latest version and restart before loading either 2.9B model. Older versions may not load or run the model correctly. Each checkpoint is a 5.84 GB BF16 diffusion model. Both use the same separate text encoder and VAE listed below. SHA-256 checksums are in releases.json. Recommended settings for standard and Turbo versions, including 2.9B, are listed below. See the Civitai release page for sample images. These are BF16 diffusion-model checkpoints. The text encoder and VAE are separate…

LoRA Flux.1-dev pour le personnage Emma Vaganova, entraînée avec le token 3mm@. Téléchargement: onglet Files and versions de ce dépôt. Utiliser 3mm@ dans le prompt (souvent en fin de phrase: Style of 3mm@). - Résolution: 896×1024 ou buckets proches (aligné sur l’entraînement multi-résolution). Les dossiers d’état d’entraînement (-state/) ne sont pas publiés ici pour l’instant: reprise Comfy = usage local uniquement. Personnage fictif inspiré de références photo; usage responsable et conforme aux conditions FLUX / OpenRAIL. Voir MODELCARD.md (steps, reprise, hyperparamètres, SHA256, limites connues).

You should use Zelretch22 to trigger the image generation. Download them in the Files & versions tab.

Model · Text to image

Qwen-Image-2.1-PE-T2I-Heretic-NVFP4

Darrellbest

NVFP4 (4-bit floating point, W4A4) build of darrellbest/Qwen-Image-2.1-PE-T2I-Heretic, the refusal-ablated text-to-image prompt rewriter for Qwen-Image-2.1. For vLLM on NVIDIA Blackwell GPUs, which run NVFP4 natively. 11 GB instead of 18 GB. systemprompt.txt is included and required, exactly as for the original. The linear-attention layers carry a recurrent state and the vision tower encodes the input image; both were left in bf16, as other quantizations of this model family do. That is why the file is 11 GB rather than ~6 GB. Made with llm-compressor 0.13.0 (QuantizationModifier, scheme="NVFP4"), calibrated on 64 samples in the model's real input format: its own system prompt and a short…

very awa, masterpiece, best quality, absurdres, perfect quality, detailed background, intricate details lowres, worst quality, bad quality, bad anatomy, signature, username, logo, watermark, artist name, bad hands You should use Zelretch22 to trigger the image generation. Download them in the Files & versions tab.

Model · Text to image

pimp-my-pony-gguf

Chuck

GGUF-quantized versions of huchukato/pimp-my-pony — an SDXL/Pony-based merge — for use with Each model variant lives in its own folder, with multiple quantization levels. More variants will be added over time. These files contain only the UNet — CLIP and VAE are not included. Requires the ComfyUI-GGUF custom node (pip install --upgrade gguf). SDXL is conv2d-heavy, so gains are smaller than on transformer models (Flux/Wan) — but still meaningful on limited VRAM / unified memory. Recommended generation settings are the same as the clip skip 2, included VAE, avoid excessive negative prompts. Converted with the official UNet extracted to safetensors → convert.py (BF16 GGUF) → llama-quantize…

Model · Text to image

agate-preview-002-4step

ML-Intern-lab

An unofficial distillation of LogoLabs' Agate Preview 002 (260M parameters, MIT) into a 4-step model with classifier-free guidance baked in: 4 network passes per image instead of the teacher's 100 (50 Euler steps × a conditional and an unconditional pass). Same architecture, same 256 px output, same agate/ code; only the weights differ. in your browser (WebGPU, nothing leaves your machine) · Use steps=4, cfg=1.0. The bundled pipeline skips the unconditional pass when cfg == 1.0. Every image carries Agate's invisible watermark (payload AGATE002) and provenance metadata; pass watermark=False only for images that go into metrics. A WebGPU build (webgpu/, same graph I/O as Agate's own) runs in…

Model · Text to image

28

Vittu

Download them in the Files & versions tab.

Model · Text to image

29

Vittu

Download them in the Files & versions tab.

Generated using https://oneshotlora.com Users are solely responsible for complying with all applicable local, national, and international laws and regulations when using this model. Do not use it for any unlawful purpose.

Context binaries (QAIRT 2.50.40, HTP v79, socmodel 69) that run circlestone-labs/Anima Turbo v1.1 entirely on the Snapdragon 8 Elite neural processing unit: prompt encoder, 28-block drawing model and picture decoder. 6 picture sizes (512 and 768 class; square, 9:16, 16:9), 4/6/8 steps, about 6–7 s per 512×512 picture at 8 steps on an iQOO 13. Optional guide picture (layout/pose control) through kohya-ss/Anima-LLLite any-test-like-v2, built into the drawing model. Non-commercial use only — see LICENSE.md and NOTICE. Generated pictures may be used commercially. This is a modified version of Anima and is not an official CircleStone Labs release. Low memory set (what the app's "Low memory"…

Animagine XL 3.1 with the SDXL-Lightning 4-step LoRA fused in, its UNet quantised to w8a16 (int8 weights, int16 activations) with Qualcomm AI Hub and compiled to a Hexagon NPU context binary for the Snapdragon 8 Gen 3 (QAIRT 2.45). 1024x1024 in 4 steps, about 11 seconds on a Snapdragon 8 Gen 3 including load. Full-size SDXL does not fit one HTP graph in fp16 (~5.5 GB against a 3.5 GB limit); int8 weights bring the UNet to 2.6 GB. The graph has fp32 IO bound by timeids -> output0. Context binaries are per chip generation. The VAE decoder comes from latentdivergence/lcm-ssd1b-npu-, the text encoders from latentdivergence/animagine-xl-lightning-int8, the tokenizer from…

Source: sca255/animemixv10lcm (SD 1.5 LCM anime model, StableDiffusionPipeline). Recipe: numinferencesteps=8, guidancescale=1.5, 1024x1024 (LCM distilled model). Converted with optimum-intel 2.2.0 / optimum 2.3.0 / openvino 2026.4.0 / nncf 3.4.0: 1. optimum-cli export openvino -m sca255/animemixv10lcm --task text-to-image --library diffusers --weight-format fp16./animemixv10lcm-ov-fp16 2. OVQuantizer NNCF weight-only INT4 (bits=4, sym=False, groupsize=128, fallback=adjust, ratio=1.0) on unet + textencoder, VAE INT8 default. See quantizeint4.py. - FP16 OV 3.2G → INT4 OV 0.86G (unet 700M, textencoder 81M, vaedec 48M, vaeenc 33M, ~73%↓). Env: diffusers==0.37.1 transformers==4.57.6…

A DreamBooth-LoRA for Krea 2, trained on Krea 2 RAW and shown on Krea 2 Turbo. The samples below were generated with this LoRA on Turbo (8 steps). Use the token vxy girl to invoke the concept.

Model · Text to image

ASSANDTITS

Wym issam Mouloungui

Download them in the Files & versions tab.

Model · Text to image

Banned-Loras

Uncle Bane

This is a collection that I have been trying to gather from LoRas that were banned or were from the very start way "too spicy" to ever upload into Civitai so hopefully users could try find some of them here for what they want to do. Im focusing mainly on LoRas for PonyXL, IllustriousXL and Anima. Because write what each LoRa respective trigger words is way too much, I uploaded a LoRa Viewer that you can use to load each of them to see what words trigger each of them. At the time of writing I can be contacted through my discord server perhaps if Im not quite busy with something (the usual) http://chat.nx34.xyz I also have another project over here: http://ubtl.nx34.xyz

Model · Text to image

Blaze5

Jacob Zoyiopoulos

It takes words and makes pictures Download them in the Files & versions tab.

Model · Text to image

Blaze55

Jacob Zoyiopoulos

It takes words and makes pictures Download them in the Files & versions tab.

You should use 5123512351251235 to trigger the image generation. Download them in the Files & versions tab.

This Hugging Face repository hosts the released LoRA checkpoints for CapField-OPD, an on-policy distillation (OPD) framework that consolidates multiple reward-specialized teachers into a single student through a continuous capability field. Existing multi-teacher OPD methods route each prompt to a single teacher according to its semantic category, implicitly binding the desired capability to prompt content. CapField-OPD instead conditions the student on an explicit capability coordinate $\boldsymbol{\lambda} \in [0,1]^K$, whose axes control the strength of each capability. The base model, single-capability teachers, and joint-capability teachers serve as anchors of the field, and their…

A DreamBooth-LoRA for Krea 2, trained on Krea 2 RAW and shown on Krea 2 Turbo. The samples below were generated with this LoRA on Turbo (8 steps). Use the token cdn3 woman to invoke the concept.

Model · Text to image

Chitra-Tiny-ONNX

Mrfakename

ONNX export of udgamlabs/Chitra-Tiny, a 212M-parameter text-to-image model (256×256), for onnxruntime-web on WebGPU. Used by the mrfakename/chitra-tiny-webgpu Space. Weights are stored in fp16 and cast to fp32 inside each graph: half the download, fp32 math. The DiT can't run in fp16 because its residual stream goes far past the fp16 range. Sampling (same as chitratiny.py): tokenize with padding to 77, encode the prompt and the negative prompt (empty by default), start from Gaussian noise, then Euler steps over t = linspace(1, 0, steps+1) with x += (t[i+1] - t[i]) (vuncond + g (vcond - vuncond)), then decode. Defaults: 50 steps, guidance 7. tools/exportonnx.py reproduces the files…

Model · Text to image

comradeshipXL

Hanzogak

Comradeship XL v7/v7A/v7PP recommend CFG - 6.5 v4BB/5A/6/6A/v12T1/v14T1S recommend CFG - 4~6 OLD Comradeship XL models recommend CFG - 2.5~3.5 This model is a merger of the Comradeship XL v14VZ model and the Experimental NoobAI with Rectified Flow + EQ-VAE model. (https://civitai.com/models/2071356?modelVersionId=2343811) This model is Rectified Flow. YOU NEED TO USE COMFYUI/SWARMUI or any webui that can do ModelSamplingSD3 at 2.5 Comradeship XL v14VW3 × WAI-NSFW-illustrious-SDXL v15 Karcher-merge V-pred model This is a minor update model of comradeshipXL-v14KC. A V-Pred merge model with slightly improved natural language prompt performance using Karcher-merge.…

Model · Text to image

coreaiimages

Privacy AI

Apple Core AI builds of open text-to-image models, mirrored for on-device generation in Privacy AI (iOS 27 / macOS 27, Core AI diffusion pipeline). Nothing here is original work: each subfolder is a redistributed model, converted with Apple's coreai.diffusion.export and, where stated, quantized. Each subfolder has its own licence. The licence text, a NOTICE, and MIRROR-NOTES.md (source repository and revision, exact export command, what was changed and dropped) sit inside every /source/ leaf, so they travel with the files to every device. Use of the Stable Diffusion v1-5 and v2-1 subfolders is subject to the use-based restrictions in Attachment A of their licences. Stable Diffusion 3.5…

Model · Text to image

Cosmos3-Super-Text2Image

Ares Realm Studios

NVIDIA Cosmos™ is a world foundation model platform designed to accelerate the development of Physical AI by enabling machines to understand, simulate, and interact with the physical world across robotics, autonomous driving, and smart space environments, including industrial and factory-scale applications. Cosmos3 is a collection of Omnimodal world models capable of generating dynamic, high-quality video, image, audio, and action commands from combinations of text, image, video, and action trajectory inputs. It serves as a foundational building block for a broad range of Physical AI applications and research spanning world understanding, world generation, simulation, and embodied policy…

Pipeline Class: CraftlyImagePipeline; Transformer Diffusion Backbone: CraftlyImageTransformer2DModel (28 Layers, 48 Attention Heads, 64-channel Latent Input); Text Conditioning Backbone: CraftlyVisionModel (36 Layers, 2560 Hidden Dim); Latent Autoencoder: 16-channel Spatial VAE; Denoising Trajectory: Flow-Matching Euler Discrete Scheduler (Optimized 8-step Distillation).

Adaptador LoRA (Low-Rank Adaptation) entrenado para generar imágenes del traje emblemático del Diablo Huma de Pujilí (patrimonio cultural inmaterial del Ecuador), como parte de un trabajo de titulación sobre preservación digital del patrimonio con IA. En este repositorio se incluye un cuaderno listo para generar imágenes con el adaptador: Cómo usarlo: descárgalo y súbelo a Google Colab (Archivo → Subir cuaderno), activa GPU (Entorno de ejecución → Cambiar tipo de entorno → T4) y ejecuta las celdas. Como alternativa confiable, el modelo también se puede usar en ComfyUI. - Palabra de activación (trigger): d14bl0huma Compatible con ComfyUI, AI-Toolkit (Ostris) y diffusers (loadloraweights).…

Powered by Stability AI. A PEFT LoRA adapter for Stable Diffusion 3.5 Medium, post-trained with DiffusionART using GenEval, CLIPScore, PickScore and HPSv2 feedback. Only the LoRA adapter is included; download the SD3.5 Medium base model separately. The adapter uses rank 32 and alpha 64. Same prompt and seed within each pair. Obtain access to the gated base model, authenticate with Hugging Face, and install the verified dependencies in an Use an official CUDA-enabled PyTorch wheel compatible with your driver. Then authenticate with an account that has access to the base model and this adapter. Download the adapter and run inference: The default settings are 512 x 512, 40 sampling steps…

Experimental ONNX conversion of DreamLite-mobile, for the independent DreamLite Lab Android test application. Original implementation: ByteVisionLab/DreamLite. Download the single shared/ package (~1.57 GB total). dreamliteweights.bin holds one exact shared copy of the large weights. Four small ONNX execution graphs select UNet/decoder and 512/1024 tensor shapes. Keep all five files in the same directory without renaming. Sizes and SHA-256 hashes are listed in shared/manifest.json. APK 0.3 defaults to 512px / 1 step and allows 512/1024 and 1/2/4-step selection in Settings. Download once; changing resolution or step count does not download another weight set. 1024px inference is native…

Model · Text to image

flowmap_grpo_sCM

Zhiqi Li

LoRA adapters (r=64, alpha=128, 12 attention/FF projections) trained with Flow-Map GRPO on the T2I-Distill MeanFlow (flux1litemeanflow.pt) and sCM (flux1litescm.pt) students of Freepik/flux.1-lite-8B, plus the online-SFT / Diffusion-NFT baselines. Code: ZhiqiLi-CG/flowmapgrpo (branch and commit in each run folder's README). Layout: checkpoints/ / / [- ]/checkpoint- /{lora, loraema, trainingstate}. lora/ is the raw learner, loraema/ the EMA adapter used for all reported numbers, trainingstate/ the full resume point. Run folders carry a README.md and the training-time fullevalmetrics.jsonl where available. Evaluation protocol for the GRPO runs: K = sampling steps - 1 (K=4 is the 5-step…

This is a PEFT LoRA derived from black-forest-labs/flux-3-image-klein-base. No validation prompt was used during training. Note: The validation settings are not necessarily the same as the training settings. You can find some example images in the following gallery: The text encoder was not trained. You may reuse the base model text encoder for inference.

This is a PEFT LoRA derived from black-forest-labs/flux-3-image-klein-base. No validation prompt was used during training. Note: The validation settings are not necessarily the same as the training settings. You can find some example images in the following gallery: The text encoder was not trained. You may reuse the base model text encoder for inference.

This is a PEFT LoRA derived from black-forest-labs/flux-3-image-klein. No validation prompt was used during training. Note: The validation settings are not necessarily the same as the training settings. You can find some example images in the following gallery: The text encoder was not trained. You may reuse the base model text encoder for inference.

This is a PEFT LoRA derived from black-forest-labs/flux-3-image-klein. No validation prompt was used during training. Note: The validation settings are not necessarily the same as the training settings. You can find some example images in the following gallery: The text encoder was not trained. You may reuse the base model text encoder for inference.

Model · Text to image

flux-3-klein-openvino-int4-awq

X Girls

Complete OpenVINO IR package for the distilled FLUX.3 Klein image-generation pipeline. The transformer uses mixed INT4/INT8 weights; the Qwen3-VL text encoder uses mixed INT4/INT8 weights; both VAE paths use INT8 weights. The tokenizer, scheduler configuration, runner, license, and sample are included in the repository. The package is self-contained: inference downloads this repository only, or accepts a local copy with --model. The complete bundle is about 4.96 GiB across 37 files. All neural-network components are OpenVINO IR. vaebnstats.npz contains the FLUX.3 VAE batch-normalization statistics required by the pipeline. - validation1024seed42.png is a fixed-seed, 1024×1024 end-to-end…

Model · Text to image

geolip-beatrix-anima

AbstractPhila

Experiments on steering Anima, a 2B illustration model built on NVIDIA Cosmos-Predict2, with a second conditioning source: Beatrix, a byte-level language model from the geolip line. Anima reads its prompt through a small language model (Qwen3 0.6B) and a light adapter, so an added signal is not drowned by a very large text encoder; that makes it a bed for testing how far a second source can steer the image without a full diffusion training run. The first experiments measure the bed itself: how its conditioning responds to mood words and to a mood direction added to it, and what LoRAs trained on the model's own mood images do. Each experiment has its own folder under experiments/ with a…

Model · Text to image

HautlenceKubera

Anshuman Singh

You should use HautlenceKubera to trigger the image generation. Download them in the Files & versions tab.

Model · Text to image

HautlenceKuberaFlux.2Klein9B

Anshuman Singh

You should use HautlenceKubera to trigger the image generation. Download them in the Files & versions tab.

Model · Text to image

Hybrid-Sensitivity-Weighted-Quantization-SDXL-ConvRot-INT8

GJL

High-fidelity ConvRot INT8 reverse hybrid quantization for SDXL diffusion models. HSWQ uses per-layer trajectory-impact measurement instead of naive uniform cast, converting only the K lowest-impact layers to ConvRot INT8 while keeping every other layer at FP16. This is highly useful for users who need to strictly manage their VRAM resources while maintaining maximum image quality. Reverse hybrid (diag → reverse): The FP16 checkpoint is the only input. A per-layer trajectory-impact measurement (sdxl/diagimpactsdxl.py) injects each candidate layer's ConvRot INT8 reconstruction into the FP16 model one at a time and runs the production sampler — recording the final-latent drift. The K…

Model · Text to image

illustrious-xl-unet-gguf

HYBRIDION

UNet only. The two text encoders (CLIP-L + OpenCLIP-G) and the VAE are not in these files — they come from any Illustrious XL / SDXL checkpoint you already worth quantizing. You keep the encoders at full precision and still drop the UNet to 1.79 GB. Quantized from illustriousXLv01.safetensors (Illustrious XL v0.1 line, modelspec.architecture: stable-diffusion-xl-v1-base). All three are 1680 tensors / 2,567,463,684 parameters. The F16 tensors that remain even in the quantized builds are the 1-D biases and the 51 4-D convolution weights — Q40/Q80 block layouts require a row length divisible by 32, and a (320, 4, 3, 3) conv does not satisfy that. Those stay exact. Pick Q80 for normal use…

Original model is here. This model created by ShoponSkp.

Model · Text to image

Image-2.1-cmf

Oleg Kirichenko

Qwen-Image-2.1 packaged as a single CMF file: the 7B image transformer, the Qwen3-VL-8B text encoder with its vision tower, and the RGBA VAE together, with the generation recipe stored inside. It runs with cortiq, a Rust engine with no Python: Apple silicon through Metal, NVIDIA GPUs through Vulkan, and a CPU fallback everywhere. Text-to-image, native transparency (RGBA), and generation from condition images. Built with Qwen. No flags needed: the file carries the recipe (1024×1024, 40 steps, no CFG). Set XDGRUNTIMEDIR=/tmp on a headless Linux box. Transparent images — ask for them in the prompt, PNG keeps the alpha channel: From condition images (repeat --image for several): sha256…

RealVisXL V5.0 by SG161222, converted to Core ML for the iPhone and iPad Neural Engine with Apple's ml-stable-diffusion, following Apple's recipe for its own iPhone SDXL: SPLITEINSUM attention, 768 x 768, recipe4.50bitmixedpalette UNet (Apple's mixed-bit palettization recipe for SDXL), UNet split into two halves, 16-bit text encoders, madebyollin/sdxl-vae-fp16-fix decoder. Used by the IMAGINE IT app. License: CreativeML Open RAIL++-M, inherited from the base model. Its use restrictions apply to anyone using these weights: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md

Model · Text to image

InGuard

Alibaba-AAIG

This repository contains the released guardrail heads for InGuard, an in-pipeline safety framework for text-to-image generation. It does not contain the five base text-to-image models. For each supported model, the release provides: - a PE-MLP classifier operating on text-encoder embeddings; and - a ConvNeXt-Base detector operating on an intermediate latent estimate. Detection step uses zero-based indexing and refers to the default deployed configuration in the InGuard code repository. The PE-MLP checkpoints contain inputdim, hiddendim, numlayers, and dropout together with modelstatedict; no separate PE-MLP config file is required. Training optimizer states have been removed from all…

InstantFusion connects heterogeneous image generators through a learned shared latent interface. Model-specific Latent AutoEncoders (LAEs) align intermediate states while preserving the information needed for reconstruction and continued denoising. The base generators remain frozen during LAE training. The shared interface supports generation acceleration, multi-preference composition, and cross-model on-policy distillation (OPD). This repository provides trained LAE interface weights. The base generators must be obtained separately. LAE reconstruction examples for Qwen-Image, FLUX.1, and SD3. Mean SSIM is reported over 100 images for each model. Intermediate states can be transferred…

Model · Text to image

iris-3b_GGUFs

Rebel AI

GGUF quantizations of SperiLabs Iris-3B, a 3B-parameter pixel-space image generator that produces RGB images directly without a VAE. By Rebel AI · ComfyUI nodes on GitHub Availability and exact filenames should be checked in this repository's Files tab. The GGUF model contains Iris diffusion weights, not the Qwen3-VL encoder. A separate text encoder is required. 1. Install ComfyUI-Iris-3B and its upstream Iris source dependencies. 2. Install city96/ComfyUI-GGUF. 3. Install the optional, isolated ggufaddon/ from the ComfyUI repository as a separate sibling custom-node folder. Do not overwrite a working W4A8/FP32 installation. 4. Copy the GGUF you want to run into…

Generated using https://oneshotlora.com Users are solely responsible for complying with all applicable local, national, and international laws and regulations when using this model. Do not use it for any unlawful purpose.

Model · Text to image

krea2-particle-sliders

Ntc

Final Boss and Eldritch, with original rank-16 LoRAs and compressed rank-8 distills. Use strength 1 for the calibrated effect. Same prompt, seed and sampler across each comparison. Built on jimmycarter/krea2-turbo-bbox, epoch-14-step-73184/transformer. These are ordinary attention LoRAs. The originals were trained directly as LoRAs; the distills compress those linear adapters. All images below are AI-generated, 768 × 768, 8 steps, guidance 0, mu=1.15. Each comparison states its seed and keeps it fixed across Original, Distill and Off. They were rendered from the released files, with no external alpha multiplier. A plain research robot becomes a heavier, battle-worn machine with a more…