Sources checked: 2026-11-19

All 26 Local AI Image Models | Features, Software & Downloads

Browse features, example outputs, supported software and download pages for anime, photography, design and image editing.

Included releases: 2026-05-19 to 2026-11-19

Hardware categories are estimates. This is not an independent speed or VRAM benchmark.

HotaruBreedAnima 2.9B v2.0

JujoHotaru

Mid-rangeAnime
Blue-haired person at an amusement park at dusk (Basic) / 2.9B v2.0 Basic / Soft / Anime
Publisher's exampleBlue-haired person at an amusement park at dusk (Basic) / 2.9B v2.0 Basic / Soft / Anime

Fine lines and soft colors for a consistent, cute look. The example shows a character at an amusement park at dusk. Choose Basic, the softer Soft, or cel-shaded Anime. The style is strongly fixed and not suited to large prompt-driven style changes.

Three variants: Basic / Soft / Anime. The example is from Basic.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://huggingface.co/JujoHotaru/BreedSeriesForAnima

AnimaIka v5.1

giko

Mid-rangeAnime
Indoor illustration of a maid petting a cat / v5.1
Publisher's exampleIndoor illustration of a maid petting a cat / v5.1

Soft anime shading, from clothing folds to indoor light. The maid-and-cat example has clear outlines and gentle shadows. v5.1 adjusts the tendency toward overly round faces while preserving body proportions.

Commercial use of generated images has not been confirmed as permitted. Check the publisher's terms.

For Anima-compatible versions of ComfyUI / Forge Neo. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2426265?modelVersionId=3467859

MiniMax-H3-Image

MiniMax

High-endPhotographyAnimeImage editingServer-class
A moonlit rooftop, a man's profile and flamingos by the water
Publisher's exampleA moonlit rooftop, a man's profile and flamingos by the water

A candidate for large images with fine cloth, hair and water detail, and the atmosphere of evening light. Supports watercolor and background concepts as well as photography, and can combine interiors or props from multiple references.

Six steps does not mean it will run in seconds on your GPU.

Server-class: about 68.65GB for the main model and related components, plus about 66.7GB for the text encoder. The standard configuration is not intended to fit on a single 5090.

Model download pagehttps://huggingface.co/MiniMaxAI/MiniMax-H3-Image

Nova Anime AM v7.0:2.9B

Crody

Mid-rangeAnime
White-haired character in purple lighting / v7.0:2.9B
Publisher's exampleWhite-haired character in purple lighting / v7.0:2.9B

Vivid light and dimensional shading between anime and 2.5D. The example bathes a character in purple and pink light. The publisher says v7.0 reduces excessive detail and improves prompt handling and detail organization.

The publisher prohibits commercial use of unedited generated images. Check the usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2604424?modelVersionId=3462792

Qwen Image 2.1 Uncensored Omni

ChatR Labs

Mid-rangeHigh-endPhotographyAnimeImage editingUncensored
Previous local output · 2026.11.18 / INT8 · 1280×720 · 35 steps · CFG 5
Previous local generationPrevious local output · 2026.11.18 / INT8 · 1280×720 · 35 steps · CFG 5

Fine outlines and soft shading suit a quiet anime scene. The example is an existing local output combining a purple nun's habit with warm morning light. Useful for describing characters, props and backgrounds together in one prompt.

Small text within images may be distorted. The example uses the INT8 version.

Lightweight formats are estimated as mid-range; standard-precision formats as high-end. ComfyUI formats are available.

Model download pagehttps://huggingface.co/chatrlabs/Qwen-Image-2.1-Uncensored-Omni-Quantized

P-H3-Image

Pruna AI

High-endPhotographyAnimeImage editing
Risograph, crayon, 1960s photography and 1970s cinema styles
Publisher's exampleRisograph, crayon, 1960s photography and 1970s cinema styles

Suited to portraits that retain fine skin texture and photos with a vintage color-film feel. Can also produce risograph and crayon styles; piling on dramatic descriptive words can push results toward stronger gloss and saturated colors.

Small faces, hands and long text may still have defects.

The publisher reports about 21GB VRAM for the Comfy 4-bit version; standard Diffusers uses about 83–101GB. Choose the appropriate distribution format.

Model download pagehttps://huggingface.co/PrunaAI/P-H3-Image

MOSS-Image SFT

OpenMOSS Team

High-endPhotographyAnimeText & design
SFT version: a person and a whiteboard with Chinese text
Publisher's exampleSFT version: a person and a whiteboard with Chinese text

Combine softly lit people and text-rich backgrounds in one image. SFT is a further-tuned version; the official comparison shows more organized whiteboard headings and paragraphs than Base. Also a candidate for posters and explanatory illustrations.

Check text accuracy after generation, especially fine Japanese text.

Categorized by standard precision. An additional integration for ComfyUI is available.

Model download pagehttps://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT

Hikari Anima v1.2.0

RedRayz

Mid-rangeAnime
Silver-haired mage with blue-violet light / v1.2.0
Publisher's exampleSilver-haired mage with blue-violet light / v1.2.0

Translucent hair, fine lines and sparkling light. The publisher's silver-haired mage example uses blue-violet illumination and iridescent reflections. For those who want to specify line and color character without relying heavily on quality tags.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2387538?modelVersionId=3408386

MOSS-Image Base

OpenMOSS Team

High-endPhotographyAnimeText & design
Base version: a person and a whiteboard with Chinese text
Publisher's exampleBase version: a person and a whiteboard with Chinese text

A base model for retaining a person's texture while directing background and text layout. Accepts natural language, tags and JSON layouts. In the official comparison, whiteboard paragraphs are less organized than in SFT, giving the versions different finishing tendencies.

If text organization matters, compare the SFT example with the same composition.

5B family. ComfyUI requires a MOSS integration. Categorized by standard precision; minimum VRAM has not been measured.

Model download pagehttps://huggingface.co/OpenMOSS-Team/MOSS-Image-Base

FLUX 3 Image Klein

Black Forest Labs

Mid-rangeHigh-endPhotographyAnimeImage editing
Official gallery; base / distilled breakdown not specified
Publisher's exampleOfficial gallery; base / distilled breakdown not specified

From naturally lit portraits to illustrations with different textures. This general-purpose model can combine several reference images or edit a specified area. The official gallery shows its range across photography, anime and text-based design.

The official gallery does not identify which examples use base or distilled.

Lightweight formats are estimated as mid-range; standard precision as high-end. Supports ComfyUI.

Model download pagehttps://huggingface.co/black-forest-labs/flux-3-image-klein

CottonAnima 2.9B_v1

uraoto

Mid-rangeAnime
Blonde character in metallic clothing / 2.9B_v1
Publisher's exampleBlonde character in metallic clothing / 2.9B_v1

Soft shading and polished anime faces for a consistent look. The example features a blonde character in metallic clothing. A merged style LoRA fixes the look, favoring consistency over broad stylistic variation.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2382223?modelVersionId=3406559

AMIXA2.9B V01

asyncforai414

Mid-rangeAnime
White-haired character in clothing with blue decorations / 2.9B_V01
Publisher's exampleWhite-haired character in clothing with blue decorations / 2.9B_V01

Clear outlines and vivid colors put the character first. The white-haired figure example has distinct color boundaries and detailed clothing. The publisher highlights Illustrious-like color rendering and clearly defined character details.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2698889?modelVersionId=3399885

Anima Cat Tower v1.1

nuko_masshigura

Mid-rangeAnime
A person sitting in tree shade beside bright grass / v1.1
Publisher's exampleA person sitting in tree shade beside bright grass / v1.1

Defined eyes and bright colors, with scene descriptions in natural language. The example shows a person in tree shade, vivid greenery and detailed clothing. v1.1 focuses on style stability with natural-language prompts.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2383017?modelVersionId=3368209

Qwen Image 2.1

Qwen

High-endPhotographyText & design
Mascot with a transparent background, displayed on white
Publisher's exampleMascot with a transparent background, displayed on white

Bright colors and clear outlines. The example mascot has a rounded face, large, neatly rendered eyes and a transparent background. Supports portraits, images with text and cutout assets.

Transparency and text quality depend on the prompt and settings.

Categorized by standard precision. Supports image generation and editing.

Model download pagehttps://huggingface.co/Qwen/Qwen-Image-2.1

Ming-Image-0.1-Design

inclusionAI / Ant Group

High-endText & designServer-class
Official Design generation example
Publisher's exampleOfficial Design generation example

A 6B model for complete app-screen concepts and text-led posters. Targets graphics with deliberate whitespace, headings and information hierarchy, and supports transparent RGBA images. More relevant to design assets than photographic portraits.

Proofread text-based designs. Lightweight versions for personal GPUs have not been verified.

Server-class: the publisher's standard BF16 validation setup uses 80GiB VRAM. Do not assume it is lightweight from the 6B name alone.

Model download pagehttps://huggingface.co/inclusionAI/Ming-Image-0.1-Design

Ming-Image-0.1-Design-Layer

inclusionAI / Ant Group

High-endText & designImage editingServer-class
Official Layer example: input design, six transparent layers and the recomposited result
Publisher's exampleOfficial Layer example: input design, six transparent layers and the recomposited result

Splits a flattened card or poster into transparent components such as background, text and decoration. Useful for turning a finished image into assets that are easier to revise, rather than choosing a new art style.

Layers are estimated by AI; this does not recover the original editable project file.

Server-class: official BF16 validation uses 80GiB VRAM. Input an image and a plan for the layers to extract.

Model download pagehttps://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer

SilvermoonMix Anima Evolved v2.3 (2.9B)

silvermoong

Mid-rangeAnime
Green-haired character on a white background / v2.3_2.9B
Publisher's exampleGreen-haired character on a white background / v2.3_2.9B

Pale colors and fine lines form a neutral base to adjust with tags. The white-background character example uses light colors and restrained shading. The publisher describes further training targeting a neutral style, steadier hands and multiple characters.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2639339?modelVersionId=3335123

JANIMA - Anima v1.0 2.9B

janxd

Mid-rangeAnime
A person surrounded by blue flowers and moonlight / v1.0 2.9B
Publisher's exampleA person surrounded by blue flowers and moonlight / v1.0 2.9B

For anime illustrations with detailed hair, eyes and backgrounds. Fine hair lines and deep blues stand out in the moonlit, flower-filled example. Also intended as a base for LoRAs without locking in one style too strongly.

Check the download page for LoRA compatibility and usage terms.

For Anima-compatible versions of ComfyUI / Forge Neo; use an up-to-date release that supports 2.9B. The mid-range category is an estimate based on size and format, not measured VRAM usage.

Model download pagehttps://civitai.com/models/2642932?modelVersionId=3242290

Anima Turbo v1.1

CircleStone Labs / Comfy Org

Mid-rangeAnime
See this version's examples in the comparison article

A distilled version for trying character-focused anime images in fewer steps. Combine tags and natural language to refine clothing, composition and lighting. Useful for exploring several directions before settling on a style.

Colors and text can change at low step counts. See the comparison article for version-labeled examples.

Supports ComfyUI. Choose the Turbo v1.1 file on the download page.

Model download pagehttps://huggingface.co/circlestone-labs/Anima

SenseNova-U1.5-8B-MoT

SenseNova

High-endPhotographyText & designImage editing
Official SenseNova U1.5 generation and editing examples
Publisher's exampleOfficial SenseNova U1.5 generation and editing examples

A general-purpose model for natural photographic light and materials, as well as text in diagrams and posters. Can edit part of a reference while retaining the person or composition. Adjust settings if details or colors become too intense.

Check text and fine details in long-text outputs or complex edits.

Publisher guidance: 24GB VRAM and 64GB RAM, with offloading. The 8B in its name is not a memory requirement.

Model download pagehttps://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT

SenseNova-U1-8B-MoT-Infographic-V3

SenseNova

High-endText & designImage editing
Official V3 example: an infographic showing baseball pitch types on an arc chart
Publisher's exampleOfficial V3 example: an infographic showing baseball pitch types on an arc chart

An infographic model for revising poster headings, rearranging body text or matching colors and style. While V2 focuses on generation, V3 also handles revising an existing infographic to a finished result.

V3 is not necessarily better than V2 for every task.

About 18B parameters in total. Supports generation and editing. V2 and V3 were trained separately from the same base.

Model download pagehttps://huggingface.co/sensenova/SenseNova-U1-8B-MoT-Infographic-V3

Anima Aesthetic v1.1

CircleStone Labs / Comfy Org

Mid-rangeAnime
See this version's examples in the comparison article

A version for finishing anime characters and costumes with more control over shading and color. Specify scenes with natural language and tags, and add character or style LoRAs. Results will not necessarily match Turbo.

Check that LoRAs match the training base and architecture. See the comparison article for examples of this version.

Supports ComfyUI. Choose the Aesthetic v1.1 file on the download page.

Model download pagehttps://huggingface.co/circlestone-labs/Anima

SenseNova-U1-8B-MoT-Infographic-V2

SenseNova

High-endText & design
Official V2 example: a comic-style fantasy-world guide infographic
Publisher's exampleOfficial V2 example: a comic-style fantasy-world guide infographic

For explanatory posters combining headings, body text, arrows and small illustrations. Prioritizes the organization and readability of text-heavy layouts over photographic atmosphere.

Small text is not guaranteed to be correct or legible.

About 18B parameters in total. Check quantization and offloading requirements. V2 focuses primarily on infographic generation.

Model download pagehttps://huggingface.co/sensenova/SenseNova-U1-8B-MoT-Infographic-V2

JoyAI-Image-Edit-Plus

Joy Future Academy / JD.com

High-endPhotographyImage editing
Official Plus example: editing reference images so a woman holds a puppy
Publisher's exampleOfficial Plus example: editing reference images so a woman holds a puppy

An editing model for combining people and objects from separate photos into one scene. An official example uses reference images of a woman and a puppy to create a picture of her holding it—combining elements beyond a global color change.

Check identity consistency and small finger shapes after editing.

For image editing. Listed by the release date of the Plus weights; the Diffusers conversion date is not treated as a new model release.

Model download pagehttps://huggingface.co/jdopensource/JoyAI-Image-Edit-Plus-Diffusers

Krea 2 Turbo

Krea AI

High-endPhotographyAnime
Official Krea example generated with Turbo
Publisher's exampleOfficial Krea example generated with Turbo

A candidate for exploring specific textures—print halftones, blur or film grain—rather than uniformly glossy photos. This 8-step generation checkpoint covers styles from photography to painting.

Requires agreement to usage terms on HF. The model card describes additional safety training.

12B. At standard precision, weights alone are about 24GB, plus other components. Check quantization and offloading requirements. Supports ComfyUI.

Model download pagehttps://huggingface.co/krea/Krea-2-Turbo

Krea 2 Raw

Krea AI

High-endPhotographyAnime
View examples for this version at the source

A foundation for training LoRAs in your own style. This non-distilled version leaves more room for adjustment than the finished look of Turbo, emphasizing further training and research rather than everyday generation.

Intended for further training and research. Start with Turbo for straightforward generation.

A non-distilled 12B model. The publisher recommends training LoRAs on Raw and generating with Turbo.

Model download pagehttps://huggingface.co/krea/Krea-2-Raw
About the categories and descriptions

This catalog includes models and new versions released between 2026-05-19 and 2026-11-19. Dates come from publisher announcements or the relevant weight-file history, not README edits. Select a release date to see its source. Dates generally follow the publisher's UTC date.

Mid-range and high-end are estimates based on model size and available formats, not quality rankings or guarantees. Quantization, resolution and offloading affect memory requirements. Server-class entries need particularly large configurations; read each model's notes.

Descriptions draw on publisher information and examples. These images were not generated under matched settings, and we have not performed a common speed or VRAM benchmark. Publisher examples and previous local outputs are labeled separately. Model licenses and image rights are separate; check both before commercial use.

The Uncensored label is used only when the publisher explicitly describes the model that way. It does not guarantee that every prompt will work. Check the distribution page for usage terms.

Editorial references: Japanese model guides by hirorohi and lilting channel. Their concrete descriptions of appearance and use cases informed the structure; their text and rankings were not copied.

hirorohi · lilting channel