# Hugging Face custom checkpoint field guide

Research snapshot: **September 21, 2026**. Recommendations reflect Hub metadata, creator cards, weight listings, release history, selected publisher images and archival discussions recorded for this snapshot. This page is not a live ranking. No checkpoint weights were downloaded and no generation benchmark was run.

The strongest first trials are **Kirazuri v4, Anima Light Lavender, Anima 2.9B, Krea2 Turbo HD and Serendipity v2**. Kroma merits attention for ongoing development. Older and experimental finds are labeled below.

A full diffusion checkpoint may still need a separate text encoder and VAE. “Merged” means trained adapters were integrated into a complete model; it does not imply all parameters were retrained. File sizes below are decimal GB, not total VRAM requirements.

Likes are cumulative. Downloads are the Hub API’s trailing-month counts, not unique users or quality scores. An API zero is reported here as unavailable/untracked rather than proof of no downloads. Repository update dates can reflect documentation or another file, not a new checkpoint release. [Hugging Face explains its counting method](https://huggingface.co/docs/hub/models-download-stats).

**1. [Kirazuri Anima v4](https://huggingface.co/motimalu/kirazuri-anima) — Overlooked artistic pick.**

Full Anima fine-tune with a documented 45,058-image curated set and high-resolution training. Explore painterly, chiaroscuro, flat-color, oil and watercolor prompts.

Hub snapshot: 2 likes; 1,642 reported monthly downloads. Repository created 2026-06-27; updated 2026-08-05.

- [split_files/diffusion_models/anima-kirazuri-v4.safetensors](https://huggingface.co/motimalu/kirazuri-anima/blob/main/split_files/diffusion_models/anima-kirazuri-v4.safetensors) — 4.18 GB.

Reuse the real official Anima encoder and VAE: this repository’s companion entries resolve to tiny LFS pointer files. Its 367 MB Turbo files are add-on LoRAs. Long descriptions help; author suggests CFG 4–5 and 30–50 steps.

[All files](https://huggingface.co/motimalu/kirazuri-anima/tree/main) · [Metadata](https://huggingface.co/api/models/motimalu/kirazuri-anima)

**2. [Anima Light Lavender](https://huggingface.co/Johnny-Z/Anima-Light-Lavender) — Fresh and appearing high in current trending results.**

Anima Base v1 post-training focused on structured natural-language descriptions up to 512 tokens. Keeps the standard approximately 2B architecture and Anima companions.

Hub snapshot: 25 likes; 772 reported monthly downloads. Repository created 2026-09-18; updated 2026-09-18.

- [anima-light-lavender.safetensors](https://huggingface.co/Johnny-Z/Anima-Light-Lavender/blob/main/anima-light-lavender.safetensors) — 4.18 GB.
- [anima-light-lavender_mxfp8.safetensors](https://huggingface.co/Johnny-Z/Anima-Light-Lavender/blob/main/anima-light-lavender_mxfp8.safetensors) — 2.84 GB.

Follow the supplied caption workflow; author recommends Euler/simple, 25 steps, CFG 4, around 1280 square, empty negative. Prompt expansion is optional. Repo created September 18.

[All files](https://huggingface.co/Johnny-Z/Anima-Light-Lavender/tree/main) · [Metadata](https://huggingface.co/api/models/Johnny-Z/Anima-Light-Lavender)

**3. [Anima 2.9B preview v1](https://huggingface.co/Gazingstars123/Anima-2.9B) — Strong adoption among newer Anima derivatives.**

Expands Anima from 28 to 40 blocks. Author reports training the inserted layers on 1.7M additional samples and a July 2026 data cutoff.

Hub snapshot: 382 likes; 36,538 reported monthly downloads. Repository created 2026-08-12; updated 2026-08-28.

- [Anima-2.9B-preview-v1.safetensors](https://huggingface.co/Gazingstars123/Anima-2.9B/blob/main/Anima-2.9B-preview-v1.safetensors) — 5.84 GB.
- [Anima-2.9B-preview-v1_int8_convrot.safetensors](https://huggingface.co/Gazingstars123/Anima-2.9B/blob/main/Anima-2.9B-preview-v1_int8_convrot.safetensors) — 3.08 GB.

Use current ComfyUI support and standard Anima encoder/VAE. Detailed prompts preferred. This preview trained new layers; it is a full deliverable checkpoint but not uniform end-to-end fine-tuning.

[All files](https://huggingface.co/Gazingstars123/Anima-2.9B/tree/main) · [Metadata](https://huggingface.co/api/models/Gazingstars123/Anima-2.9B)

**4. [Krea2 Turbo HD V1](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1) — Hugging Face official-release-channel discovery.**

Custom Krea2 image weights and a jointly tuned HD VAE, targeting texture and clarity. The creator identifies Hugging Face as the only official release channel.

Hub snapshot: 101 likes; 28,780 reported monthly downloads. Repository created 2026-07-05; updated 2026-07-13.

- [Krea2-Turbo-HD-V1-Q5_K_S.gguf](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1/blob/main/Krea2-Turbo-HD-V1-Q5_K_S.gguf) — 9.51 GB.
- [Krea2-Turbo-HD-V1-fp8mixed.safetensors](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1/blob/main/Krea2-Turbo-HD-V1-fp8mixed.safetensors) — 14.13 GB.
- [Krea2-HD-vae.safetensors](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1/blob/main/Krea2-HD-vae.safetensors) — 0.51 GB.

Use its matching HD VAE. The current tree distributes quantizations of the creator’s fine-tune; there is no BF16 transformer file. Weight-file size is not total runtime VRAM.

[All files](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1/tree/main) · [Metadata](https://huggingface.co/api/models/wikeeyang/Krea2-Turbo-HD-V1)

**5. [Serendipity V2](https://huggingface.co/Raxephion/Krea2-Serendipity-V2) — Overlooked cinematic photography pick.**

Camera placement, environmental depth and multiple-subject staging are the stated priorities. The author trained an adapter on hand-captioned studio photographs and merged it into official Krea2 Turbo.

Hub snapshot: 18 likes; 4,797 reported monthly downloads. Repository created 2026-08-18; updated 2026-09-03.

- [Serendipity-V2-Fp8.safetensors](https://huggingface.co/Raxephion/Krea2-Serendipity-V2/blob/main/Serendipity-V2-Fp8.safetensors) — 13.14 GB.
- [Serendipity-V2-Bf16.safetensors](https://huggingface.co/Raxephion/Krea2-Serendipity-V2/blob/main/Serendipity-V2-Bf16.safetensors) — 25.64 GB.

A complete merged diffusion checkpoint. Reuse Krea2 encoder/VAE and describe camera position, framing and light. Published warm backlit portrait example was visually inspected; broader claims remain unbenchmarked.

[All files](https://huggingface.co/Raxephion/Krea2-Serendipity-V2/tree/main) · [Metadata](https://huggingface.co/api/models/Raxephion/Krea2-Serendipity-V2)

**6. [Kroma v0.3 Turbo](https://huggingface.co/lodestones/Kroma) — Active custom Krea2 development.**

Continued full fine-tuning, with full checkpoint distribution from v0.2 onward and a Turbo delta merged for shorter sampling.

Hub snapshot: 337 likes; downloads unavailable/untracked. Repository created 2026-07-31; updated 2026-09-20.

- [kroma-v0.3-turbo.safetensors](https://huggingface.co/lodestones/Kroma/blob/main/kroma-v0.3-turbo.safetensors) — 25.64 GB.
- [kroma-v0.3-base.safetensors](https://huggingface.co/lodestones/Kroma/blob/main/kroma-v0.3-base.safetensors) — 51.28 GB.

README still covers v0.2. Commit history dates v0.3 base to August 19 and Turbo to August 30; September 20 changes concern specialized experimental weights. Downloads are explicitly untracked on its page. Original files are large.

[All files](https://huggingface.co/lodestones/Kroma/tree/main) · [Metadata](https://huggingface.co/api/models/lodestones/Kroma)

**7. [Flux2 Klein 9B True V3](https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V3) — Strong adoption and photographic refinement.**

Full custom Klein fine-tune for photographic texture and composition, also demonstrated for image editing. Author comparison images show less smooth-looking skin; this is curated evidence.

Hub snapshot: 157 likes; 55,761 reported monthly downloads. Repository created 2026-06-25; updated 2026-07-13.

- [Flux2-Klein-9B-True-V3-Q5_K.gguf](https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V3/blob/main/Flux2-Klein-9B-True-V3-Q5_K.gguf) — 6.81 GB.
- [Flux2-Klein-9B-True-V3-fp8mixed.safetensors](https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V3/blob/main/Flux2-Klein-9B-True-V3-fp8mixed.safetensors) — 9.43 GB.

Use the Klein 9B workflow and companions. The card retains the base’s noncommercial terms despite conflicting Apache metadata. Quantized weights still need encoder, VAE and activation memory.

[All files](https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V3/tree/main) · [Metadata](https://huggingface.co/api/models/wikeeyang/Flux2-Klein-9B-True-V3)

**8. [Soliloquy V2](https://huggingface.co/Raxephion/Krea2-Soliloquy-V2) — Fresh, low-visibility photographic atmosphere.**

Full merged checkpoint for cinematic portraits, tactile surfaces and photographic surrealism. Its trained photographic adapter is merged into VINCE1968’s Badmilkmelancholy Krea2 variant.

Hub snapshot: 2 likes; 333 reported monthly downloads. Repository created 2026-09-03; updated 2026-09-17.

- [Soliloquy-V2-Fp8.safetensors](https://huggingface.co/Raxephion/Krea2-Soliloquy-V2/blob/main/Soliloquy-V2-Fp8.safetensors) — 12.82 GB.

Its lineage differs from Serendipity. BF16 is described as a reconstruction, not an original higher-precision training export. Published changes target gentler tonal transitions and more selective detail.

[All files](https://huggingface.co/Raxephion/Krea2-Soliloquy-V2/tree/main) · [Metadata](https://huggingface.co/api/models/Raxephion/Krea2-Soliloquy-V2)

**9. [ERNIE Image Flex V1](https://huggingface.co/wikeeyang/Ernie-Image-Flex-V1) — Overlooked alternative foundation.**

Full ERNIE-Image fine-tune for cleaner colors and materials. Its sample sheet spans Chinese poster layouts, products, food, interiors, photography and illustration.

Hub snapshot: 9 likes; 2,095 reported monthly downloads. Repository created 2026-05-11; updated 2026-05-12.

- [Ernie-Image-Flex-V1-fp8mixed.safetensors](https://huggingface.co/wikeeyang/Ernie-Image-Flex-V1/blob/main/Ernie-Image-Flex-V1-fp8mixed.safetensors) — 8.65 GB.
- [Ernie-Image-Flex-V1-bf16.safetensors](https://huggingface.co/wikeeyang/Ernie-Image-Flex-V1/blob/main/Ernie-Image-Flex-V1-bf16.safetensors) — 16.07 GB.

Author recommends CFG 1 and 16–25 steps; optional Turbo adapter shortens this. Includes workflow. A little text accuracy is traded away. This is a May release with a small audience, not a September breakout.

[All files](https://huggingface.co/wikeeyang/Ernie-Image-Flex-V1/tree/main) · [Metadata](https://huggingface.co/api/models/wikeeyang/Ernie-Image-Flex-V1)

**10. [AAAAnima](https://huggingface.co/Anzhc/AAAAnima) — Unusual artistic experiment.**

Anima adaptation for 2.5D imagery, abstract scenes, darker atmosphere and horror. The later version experiments with several Qwen text-encoder layers.

Hub snapshot: 11 likes; 122 reported monthly downloads. Repository created 2026-06-01; updated 2026-08-19.

- [AAAAnima not-so-early.safetensors](https://huggingface.co/Anzhc/AAAAnima/blob/main/AAAAnima%20not-so-early.safetensors) — 4.18 GB.
- [AAAAnima Multi-layer text proj.safetensors](https://huggingface.co/Anzhc/AAAAnima/blob/main/AAAAnima%20Multi-layer%20text%20proj.safetensors) — 4.19 GB.

The ordinary file is the simpler entry point. Multi-layer text projection requires the author’s custom node. Some character knowledge may be lost. Claims concern author tests and training intent.

[All files](https://huggingface.co/Anzhc/AAAAnima/tree/main) · [Metadata](https://huggingface.co/api/models/Anzhc/AAAAnima)

**11. [AnimaYume v1.5](https://huggingface.co/duongve/AnimaYume) — Recent character and concept expansion.**

Author reports finishing training September 5, with newer anime/furry concepts and Summer 2026 characters in an ordinary approximately 2B footprint.

Hub snapshot: 57 likes; downloads unavailable/untracked. Repository created 2026-03-08; updated 2026-09-06.

- [split_files/diffusion_models/AnimaYume_v15_base.safetensors](https://huggingface.co/duongve/AnimaYume/blob/main/split_files/diffusion_models/AnimaYume_v15_base.safetensors) — 4.18 GB.

Explicitly a base without aesthetic polishing. Long prompts help; the author warns that Anima-v1 LoRAs may not transfer reliably. Repository age predates this version.

[All files](https://huggingface.co/duongve/AnimaYume/tree/main) · [Metadata](https://huggingface.co/api/models/duongve/AnimaYume)

**12. [SilvermoonMix Anima Evolved v2.3 2.9B](https://huggingface.co/silvermoong/SilvermoonMix-Anima-Evolved) — Very recent, almost invisible Hub presence.**

Author-branded checkpoint family targeting neutral illustration, artist tags and interactions between characters. The Hub repository was created September 11 and updated September 19.

Hub snapshot: 0 likes; 221 reported monthly downloads. Repository created 2026-09-11; updated 2026-09-19.

- [SilvermoonMix-Anima29B-Evolved-V2.3.safetensors](https://huggingface.co/silvermoong/SilvermoonMix-Anima-Evolved/blob/main/SilvermoonMix-Anima29B-Evolved-V2.3.safetensors) — 5.84 GB.

Use current Anima 2.9B support and standard companions. Standard and Turbo versions have different sampling settings. Links back to Civitai; no removal evidence.

[All files](https://huggingface.co/silvermoong/SilvermoonMix-Anima-Evolved/tree/main) · [Metadata](https://huggingface.co/api/models/silvermoong/SilvermoonMix-Anima-Evolved)

**13. [AniSee v2](https://huggingface.co/SeeSee21/AniSee) — Convenient clean anime variants.**

Base, Aesthetic and Turbo variants ship as full diffusion weights. The publisher includes same-seed comparisons and workflows.

Hub snapshot: 24 likes; 329 reported monthly downloads. Repository created 2026-05-17; updated 2026-09-05.

- [AniSee-V2-Aesthetic.safetensors](https://huggingface.co/SeeSee21/AniSee/blob/main/AniSee-V2-Aesthetic.safetensors) — 4.18 GB.
- [AniSee-V2-Turbo.safetensors](https://huggingface.co/SeeSee21/AniSee/blob/main/AniSee-V2-Turbo.safetensors) — 4.18 GB.

V2 is a trained LoKr baked into full weights; v1 used full fine-tuning. Reuse Anima companions. Turbo example: 12 steps and CFG 1; Aesthetic: around 40 steps and CFG 4–5.

[All files](https://huggingface.co/SeeSee21/AniSee/tree/main) · [Metadata](https://huggingface.co/api/models/SeeSee21/AniSee)

**14. [Beyond Reality Z Image v3](https://huggingface.co/Nurburgring/BEYOND_REALITY_Z_IMAGE) — Established Z-Image portrait alternative.**

LoRA training and merging produce a complete Z-Image Turbo model targeting film-like portrait tonality, fabric/skin texture and light.

Hub snapshot: 164 likes; 3,839 reported monthly downloads. Repository created 2025-12-21; updated 2026-02-10.

- [BEYOND REALITY SUPER Z IMAGE 3.0 浓妆淡抹总相宜 FP8.safetensors](https://huggingface.co/Nurburgring/BEYOND_REALITY_Z_IMAGE/blob/main/BEYOND%20REALITY%20SUPER%20Z%20IMAGE%203.0%20%E6%B5%93%E5%A6%86%E6%B7%A1%E6%8A%B9%E6%80%BB%E7%9B%B8%E5%AE%9C%20FP8.safetensors) — 6.15 GB.

Author suggests Euler/simple, CFG 1, 10–15 steps. The oddly named 19 GB ZIP/safetensors item is an archive, not the image checkpoint. Updated February: an established find.

[All files](https://huggingface.co/Nurburgring/BEYOND_REALITY_Z_IMAGE/tree/main) · [Metadata](https://huggingface.co/api/models/Nurburgring/BEYOND_REALITY_Z_IMAGE)

**15. [Juggernaut Z Image](https://huggingface.co/RunDiffusion/Juggernaut-Z-Image) — Familiar trainer on a less obvious foundation.**

Full Z-Image Base fine-tune with published comparisons for cinematic lighting, portraits, editorial work and anatomy.

Hub snapshot: 132 likes; 5,791 reported monthly downloads. Repository created 2026-04-29; updated 2026-05-13.

- [Juggernaut_Z_V1_FP8_e4m3fn.safetensors](https://huggingface.co/RunDiffusion/Juggernaut-Z-Image/blob/main/Juggernaut_Z_V1_FP8_e4m3fn.safetensors) — 6.15 GB.

This uses Z-Image Base sampling: the author’s default is roughly 35 steps, CFG 6. Do not apply stock Turbo settings. A May-era alternative rather than a new September model.

[All files](https://huggingface.co/RunDiffusion/Juggernaut-Z-Image/tree/main) · [Metadata](https://huggingface.co/api/models/RunDiffusion/Juggernaut-Z-Image)

**16. [Real Qwen Image v2](https://huggingface.co/wikeeyang/Real-Qwen-Image-V2) — Custom Qwen realism option.**

A real Qwen-Image-2512 fine-tune emphasizing photographic sharpness and Asian portrait aesthetics while retaining text generation.

Hub snapshot: 44 likes; 1,818 reported monthly downloads. Repository created 2026-01-17; updated 2026-01-18.

- [Real-Qwen-Image-V2-2512-fp8_e4m3fn.safetensors](https://huggingface.co/wikeeyang/Real-Qwen-Image-V2/blob/main/Real-Qwen-Image-V2-2512-fp8_e4m3fn.safetensors) — 20.43 GB.

Large weights and companion memory make this a heavier local experiment. Q4 and alternative quantized variants exist. January release; included for distinctive lineage, not freshness.

[All files](https://huggingface.co/wikeeyang/Real-Qwen-Image-V2/tree/main) · [Metadata](https://huggingface.co/api/models/wikeeyang/Real-Qwen-Image-V2)

**17. [Z Image Turbo Art](https://huggingface.co/wikeeyang/Z-Image-Turbo-Art) — Decorative art and advertising direction.**

Layer fusion of Z-Image Base and Turbo, followed by fine-tuning, emphasizes stylization and decorative imagery.

Hub snapshot: 48 likes; 255 reported monthly downloads. Repository created 2026-01-30; updated 2026-02-01.

- [Z-Image-Turbo-Art-Q8_0.gguf](https://huggingface.co/wikeeyang/Z-Image-Turbo-Art/blob/main/Z-Image-Turbo-Art-Q8_0.gguf) — 7.22 GB.

Author explicitly trades some delicate portrait rendering for stronger stylization. Suggested CFG 1 and 8–12 steps. February-era release.

[All files](https://huggingface.co/wikeeyang/Z-Image-Turbo-Art/tree/main) · [Metadata](https://huggingface.co/api/models/wikeeyang/Z-Image-Turbo-Art)

**Further experiments worth knowing about**

- [Anima 3.8B v1.1](https://huggingface.co/lylogummy/Anima-3.8B): 113 likes; August release. Expanded blocks plus a semantic connector and Qwen3.5 4B conditioning target spatial relationships and several characters. Requires custom nodes and dual encoders. Current full file is `difussion_models/Anima-3.8B-v1.1.safetensors` (8.81 GB); the directory spelling is real. Heavier than ordinary Anima.
- [MageTrail](https://huggingface.co/RicemanT/MageTrail): 17 likes; September 4–21 repository activity. An alternative MageFlow-based anime/furry training project. The author calls the documented v0.1/v0.2 runs undertrained. New v0.3 epoch files are 5.97 GB while the README still describes earlier 8.23 GB files; do not assume the same architecture or loading requirements.
- [AAAAnima multi-layer projection](https://huggingface.co/Anzhc/AAAAnima) is the unusual artistic experiment in the main list. [Zeta-Chroma](https://huggingface.co/lodestones/Zeta-Chroma) and [Chroma2 Kaleidoscope](https://huggingface.co/lodestones/Chroma2-Kaleidoscope) also have full weights, but sparse WIP documentation does not support a polished-quality recommendation.
- [Pattern Diffusion](https://huggingface.co/Arrexel/pattern-diffusion): a distinct older specialist, trained from scratch on seamless patterns using an SD2-style architecture. Useful for textiles and surface designs. This is a specialist foundation, not a recent custom aesthetic merge; its inference recipe uses noise rolling and late circular padding.

**Specific recoveries and archive discoveries**

**[ORPHIC 3](https://huggingface.co/John6666/orphic-3-sdxl)** — SDXL/Pony anime/realism crossover. Ungated `orphic_3.safetensors` (6.94 GB) and Diffusers files remain available. A user reported Civitai removal in October 2024 and the mirror maintainer replied that they had restored the original file from disk. This is documented historical recovery; no creator explanation for removal was found. [Preservation discussion](https://huggingface.co/John6666/orphic-3-sdxl/discussions/1).

**[Beautiful Art v4–v7](https://huggingface.co/LMFResearchSociety/CheckpointArchive/tree/main/lusciousfruits)** — legacy SD1.x artistic checkpoints. Files `beautifulArt_v40/v50/v60/v70.safetensors` are listed at approximately 2.13/2.13/2.74/2.13 GB. Access is manually gated. A merged archivist contribution explicitly records Civitai removal. [Discussion](https://huggingface.co/LMFResearchSociety/CheckpointArchive/discussions/6) and [discussion API](https://huggingface.co/api/models/LMFResearchSociety/CheckpointArchive/discussions/6). [Yntec’s BeautifulArt](https://huggingface.co/Yntec/BeautifulArt) is an ungated related v1 rescue, not the same v4–v7 weights.

**[AmalgamMix](https://huggingface.co/LMFResearchSociety/CheckpointArchive/blob/main/PloughB660/amalgamMix_.safetensors)** — 2.13 GB legacy merge, manually gated. [Jemnite’s experiment notes](https://huggingface.co/Jemnite/experiments) explicitly report deletion of its Civitai page and identify the surviving archives. Interesting for older stylized illustration recipes; some original merge settings were lost.

**[Andite’s artist-checkpoint backup](https://huggingface.co/LMFResearchSociety/andite-finetunes-backup)** — ungated full WLOP, GUWEIZ and Yohan models, plus recovered Dessert variants such as Tea. The [archivist’s account](https://huggingface.co/LMFResearchSociety/andite-finetunes-backup/discussions/1) concerns deletion of Andite’s Hugging Face account. Civitai removal is not established. These are old artistic models, included for recovery value.

**[CinEro-Krea2-Dark v1e All-In-One](https://huggingface.co/Quiho/civitai-archive/blob/main/cineroKrea2Dark_v1eAllinone.safetensors)** — a newer niche archive find by homoludens, published August 30, 2026. Dark cinema, sci-fi/horror and atmospheric landscapes; 31.14 GB including encoder and VAE. Hub-recorded SHA256 matches the [preserved Civitai metadata](https://huggingface.co/Quiho/civitai-archive/blob/main/cineroKrea2Dark_v1eAllinone.json). [Preserved creator description](https://huggingface.co/Quiho/civitai-archive/blob/main/cineroKrea2Dark_v1eAllinone.html). Mirror provenance is verified; deletion is not. Newer versions exist elsewhere.

**Productive creator pages and archive collections**

- [wikeeyang](https://huggingface.co/wikeeyang): several original image-model families, typically with workflows, comparison images and quantizations.
- [Raxephion](https://huggingface.co/Raxephion): Serendipity, Soliloquy, Solstice and Ronin. Ronin is built on an abandoned community checkpoint with incomplete ancestry; abandonment does not establish deletion.
- [EllaPriest45/Krea2_Checkpoints](https://huggingface.co/EllaPriest45/Krea2_Checkpoints): backups including Amazing Reality, BF95 Dark Realism v3, Dark Beast v3, Dirty Realism v3, FinalCut and Jib Mix Jalapeno. The preservation purpose is explicit; specific deletions are unproven.
- [Quiho/civitai-archive](https://huggingface.co/Quiho/civitai-archive): newer full checkpoints plus preserved metadata. Collection counters cannot be assigned to an individual checkpoint.
- [Chroma111/CivitAI-Archive](https://huggingface.co/Chroma111/CivitAI-Archive) and [Archive-2](https://huggingface.co/Chroma111/CivitAI-Archive-2): large mixed archives; inspect model type and weight size per item.
- [LMFResearchSociety/CheckpointArchive](https://huggingface.co/LMFResearchSociety/CheckpointArchive): older recoveries, currently manually gated. The current card warns that storage pressure could lead to future deletion.

**Attractive search results that were not full checkpoints**

[girlslike-krea2](https://huggingface.co/ifmylove2011/girlslike-krea2), [Krea2-realism-V2](https://huggingface.co/RudySen/Krea2-realism-V2), [M87](https://huggingface.co/mgwr/M87), [Krea2 Turbo SDA](https://huggingface.co/F16/krea2-turbo-sda), [Krea2 Pixel Art](https://huggingface.co/e-n-v-y/Krea-2-Pixel-Art) and [Krea2 Lineart Edit](https://huggingface.co/Cierpliwy/krea2_lineart_edit) are adapters. They can still be useful, but their popularity should not be mistaken for adoption of distinct full checkpoints. BennyDaBall’s Engineer releases in this search are text encoders.

**Suggested starting points**

For artistic illustration, begin with Kirazuri or Light Lavender, then compare Anima 2.9B. For photography, consider Flux True V3, Krea2 HD or Serendipity. Check each model’s workflow, companion components and memory needs before choosing a precision. File size alone does not determine runtime memory use. Krea2 HD requires its matching HD VAE.

Publisher images inspected include [Kirazuri’s example](https://huggingface.co/motimalu/kirazuri-anima/resolve/main/example.png), [Serendipity’s sample](https://huggingface.co/Raxephion/Krea2-Serendipity-V2/resolve/main/images/01.png), [Krea2 HD’s sample](https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1/resolve/main/images/001.jpg), [Flux True V3’s comparison](https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V3/resolve/main/2-Compare-with-offical-and-old-version.jpg) and [ERNIE Flex’s sample sheet](https://huggingface.co/wikeeyang/Ernie-Image-Flex-V1/resolve/main/v1-sample.jpg). These are creator-selected examples, not images generated during this research.


**Additional catalog navigation**

- [Solstice V2](https://huggingface.co/Raxephion/Krea2-Solstice-V2): full merged Krea2 weights for cinematic fantasy and character photography.
- [Ronin V1](https://huggingface.co/Raxephion/Krea2-Ronin-V1): full merged weights built on an abandoned community base; complete ancestry is unresolved, and abandonment does not prove deletion.
- [Jemnite’s experiment collection](https://huggingface.co/Jemnite/experiments): manually gated legacy merge recipes and preservation context.
- [Z-Image Engineer V6](https://huggingface.co/BennyDaBall/Z-Image-Engineer-V6) and [Krea2 Engineer V1](https://huggingface.co/BennyDaBall/Krea-2-Engineer-V1): text encoders, not image diffusion checkpoints.

All external references from the source research guide are retained in the catalog’s searchable resource directory. Gating, availability, repository dates and popularity metrics describe the September 21, 2026 snapshot.