HexAnima 1.4.3

HexAnima 1.4.3

Preview at every step. While a picture is made, the main screen shows what it looks like now and sharpens it after every step (all models). It is a quick approximation, not a full decode, so it costs no time.

Qwen-Image takes up to three pictures ("the girl from picture 1 stands in front of the cafe from picture 2"): add more pictures under the photo with "+ Another picture" and refer to them as "picture 1/2/3" in the prompt. Each picture is reduced so that all of them fit (about 0.3 Mpx each for two, 0.2 Mpx for three). Two or three pictures take about the same time as one edit.

No more crashes on low memory. When the phone starts to thrash, the generation now stops with a warning instead of Android closing the app. Close other apps or pick a smaller size and try again. A hint also appears at the start when little memory is free.

History: filter by model (chips under the tabs). Main screen: tapping a picture shows "Reuse settings" there too. Temperature: the battery temperature and the distance to throttling are shown under the RAM strip.

Install over 1.4.2, 1.4.1, 1.4.0 or 1.3.0: models, history and settings stay. Check the file with sha256sum -c SHA256SUMS.txt.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). Other chips are not supported. Tested on a 12 GB phone only; on 8 GB phones big Qwen edits (about 1 Mpx) can still run out of memory, the app now warns instead of closing.

Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba (Qwen), Microsoft, LightX2V, OpenMOSS, Unsloth, Black Forest Labs or the model authors.

HexAnima 1.4.2

HexAnima 1.4.2

Your own LoRA for the Anima models (3D Harem, RealSkin; 2B and 2.9B). Pick a .safetensors file in the new LoRA block on the main screen: the app checks it at once (kohya, diffusers/PEFT, LyCORIS LoKr and LoHa; block count, rank, layer sizes, data type) and says why if it does not fit. A strength slider sets how strong it is. The LoRA is folded into the model weights when the model loads: the first picture with a new LoRA (or strength) takes 10-15 seconds longer, then it is cached; the folded weights are kept as 8-bit blocks (about 2 GB of storage for the last LoRA). LoRAs for the 2B model also work on the 2.9B models and back. The text-encoder part of a LoRA is not applied. The LoRA used is saved in the picture's details.

The native folding code is based on a pull request by GrimmDev-Glitch, thank you!

Install over 1.4.1, 1.4.0 or 1.3.0: models, history and settings stay. Check the file with sha256sum -c SHA256SUMS.txt.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). Other chips are not supported.

Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba (Qwen), Microsoft, LightX2V, OpenMOSS, Unsloth, Black Forest Labs or the model authors.

HexAnima 1.4.1

HexAnima 1.4.1

Mage-Flow-Edit-Turbo support (Microsoft, 4B, MIT license): text-to-image and editing a photo by a prompt in 4 steps, about 12 seconds a picture (about 20 s for an edit). The text encoder (Qwen3-VL-4B), its vision tower, the 12-block transformer and the VAE run on the NPU. The ~8.7 GB pack is published on Hugging Face: open the model manager and download "Mage-Flow-Edit-Turbo" (it needs this version, 1.4.1 or newer). Pictures go up to 6144 patches (e.g. 1536x1024), edits up to 3200 (1024x768). Examples are in the README.

Also:

• Fixed a crash when a generation failed right after it started (the app was killed with "did not call startForeground"): the foreground service is now stopped through itself. It affected every model.
• The model catalog is now read from catalog3.json first (then catalog2.json, catalog.json), so Mage does not show up in older versions that could not run it.
• Licenses: the MIT text and the changes made to Mage-Flow are in Info -> Licenses; THIRD_PARTY and the README list the Microsoft and Qwen3-VL components. The new texts are translated into all languages of the app. The log report now includes mage.log.

Install over 1.4.0 or 1.3.0: models, history and settings stay. Check the file with sha256sum -c SHA256SUMS.txt.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). Other chips are not supported. Mage-Flow is tested on a 12 GB phone only.

Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba (Qwen), Microsoft, LightX2V, OpenMOSS, Unsloth, Black Forest Labs or the model authors.

HexAnima 1.4.0

HexAnima 1.4.0

Qwen-Image 2.1 Turbo support (Qwen team, 4-step Lightning adapter): legible text in pictures, long prompts, and editing a photo by a prompt ("make the sky a sunset"). Text encoder (Qwen3-VL-8B), vision tower, 20B transformer and VAE run on the NPU; about 40 s for 512x768 and 40-70 s for an edit. Non-commercial use only (Qwen Research License, "Built with Qwen"). The ~14.9 GB pack is now published on Hugging Face: open the model manager and download "Qwen-Image 2.1 Turbo" (it needs this version, 1.4.0 or newer; older versions cannot run it). Examples are in the README.

For 8 GB phones (issue #1): new low-memory mode for the Anima models (Settings; on by default below ~9.5 GB of RAM). The DiT blocks are streamed from storage through two buffers instead of being pinned in RAM: about 2.5 GB less RAM for the 2.9B models, about +1 s per 512x768 picture. Not yet tested on a real 8 GB phone, feedback welcome.

MOSS-Image: removed the faint 4-pixel lattice (dotted/grid noise in dark fur, sand, hair) that the 4-step adapter leaves; on the reported seed its strength in the fur dropped about 15x. The decoder itself was checked against the CPU and is exact, so the fix is a small filter after decoding (+0.4-1 s).

Also:

• Built-in picture viewer: tap a picture (History or the main screen) to open it full screen; swipe left/right between pictures, pinch / double-tap to zoom, same actions as under a fresh picture (download, share, favourite, reuse settings, trash).
• The RAM strip now also shows how much the app itself holds (its processes and the pinned model weights), e.g. "app 1.8 GB".
• Fixed coloured dots at the seams of the Qwen-Image decoder tiles.
• The time estimate for photo editing is closer to reality; the new texts are translated into all languages of the app; THIRD_PARTY / licenses list the Qwen components. Sources of the Qwen-Image conversion are in qwen/.

Install over 1.3.0: models, history and settings stay. Check the file with sha256sum -c SHA256SUMS.txt.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). Other chips are not supported.

Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba (Qwen), LightX2V, OpenMOSS, Unsloth, Black Forest Labs or the model authors.

HexAnima 1.3.0

HexAnima 1.3.0

New: MOSS-Image 5B, a second, larger model (OpenMOSS, Apache-2.0). Describe pictures in plain sentences instead of tags. The text encoder (Qwen3.5-9B), the transformer and the decoder all run on the NPU.

• About 30 s for 512x768, about 60 s for 768x1024 and about 80 s for 1 Mpx on a Snapdragon 8+ Gen 1 (4 steps). Any size with sides that are multiples of 16 works (at least 256, at most 6144 patches).
• Optional download of about 14 GB (Models -> MOSS-Image 5B, resumable, sha256-checked). It was tested on a 12 GB phone only; the weights are streamed from storage for every picture.
• Text in the picture often works for short words but can be garbled for longer or secondary text; see the README for four unedited examples.
• MOSS-Image only does text-to-image (no image editing).

New: your own LoRA for MOSS-Image. Pick a .safetensors file: the app checks it at once (layer names, shapes, rank, data type) and tells you why if it does not fit (FLUX/SDXL LoRAs, wrong shapes, rank above 128, text-encoder-only files). A strength slider goes from 0 to 2. The LoRA is folded with the built-in 4-step adapter on the phone (about 15 s the first time for a new LoRA or strength) and never leaves the device. LoRAs stay under their authors' licenses.

Also:

• Quality tags are added and shown in Settings for the Anima models only.
• Info explains MOSS-Image, sizes, LoRA and licenses; the Apache License 2.0 text and the MOSS-Image notice (changes made to the converted files) are in Info -> Licenses.
• Fixed a rounding error in the 8-bit weight dequantisation on Hexagon V69 (some weights were one quantisation level off): the 2.9B Anima models are slightly more accurate.
• The new texts are translated into all languages of the app.
• The model catalog is now catalog2.json; older apps keep reading catalog.json and do not see MOSS-Image.

Install over 1.2.2: models, history and settings stay. Check the file with sha256sum -c SHA256SUMS.txt.
Sources of the MOSS-Image pack tools, DSP kernels and runner are in moss/.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). Other chips are not supported.

Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba (Qwen), OpenMOSS, Unsloth, Black Forest Labs or the model authors.

HexAnima 1.2.2

HexAnima 1.2.2

Snapdragon 8 Elite (Hexagon V79, Galaxy S25 and similar) support is withdrawn. It was experimental in 1.2.0 and 1.2.1. Testing on a Galaxy S25 showed that it is not reliable with the current NPU kernels:

• after an interrupted run the NPU matrix unit is sometimes silently unavailable, and pictures come out as colourful blocks (a one-tile self-test returns 0 instead of 32);
• large sizes (about 4800 tokens and more, for example 960x1536) come out as colourful noise on top of a correct picture, even when the matrix unit works. The same sizes are clean on Snapdragon 8 Gen 1 / 8+ Gen 1.

The app now reports 8 Elite phones as not supported again, the V79 files are no longer shipped (the APK is 3 MB smaller), and the README says so. If you were on 1.2.0 or 1.2.1 with an 8 Elite phone, there is nothing to do except to uninstall the app. Nothing changes for Snapdragon 8 Gen 1 / 8+ Gen 1 phones, you can install over 1.2.1 and keep your models and history.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475). A port to other chips (8 Gen 2, 8 Gen 3, 8 Elite) is not planned in the short term.

Install: download HexAnima-1.2.2.apk, allow installs from unknown sources, open it. Check with sha256sum -c SHA256SUMS.txt.
Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba or the model authors.

HexAnima 1.2.1

Superseded by 1.2.2. The Snapdragon 8 Elite support described below turned out to be unreliable (garbage pictures after interrupted runs, noise at large sizes) and was withdrawn. Snapdragon 8 Gen 1 / 8+ Gen 1 phones are not affected.

Known issue on Snapdragon 8 Elite (not on 8 Gen 1 / 8+ Gen 1): some pictures can come out as colourful noise or blocks. A tester sees it intermittently; I am still investigating (it is not tied to the Stop button, as I first suspected). If it happens, generate again or restart the app. 8 Elite support stays experimental.

HexAnima 1.2.1

Snapdragon 8 Elite (Hexagon V79) is about 4 times faster now. On a Galaxy S25 a 2.9B picture at 512x768 takes about 6 s (about 10 s with the first model load) instead of about 23 s in 1.2.0. That is faster than the 12 s of a Snapdragon 8+ Gen 1.

Cause: the weight copy asked the memory prefetcher for 4 KB in a form that Hexagon V79 does not accept, so it ran at a thirtieth of its speed. It is now requested in a form that both V69 and V79 handle well. Snapdragon 8 Gen 1 / 8+ Gen 1 speed is unchanged.

Snapdragon 8 Elite stays marked experimental: it has been tested on one phone (a Galaxy S25). Reports from other 8 Elite phones are welcome.

If you are on 1.2.0, just install this APK over it: models and history stay.

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475, tested); Snapdragon 8 Elite (SM8750, experimental). Snapdragon 8 Gen 2 and 8 Gen 3 are not supported yet.

Install: download HexAnima-1.2.1.apk, allow installs from unknown sources, open it. Check with sha256sum -c SHA256SUMS.txt.
Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba or the model authors.

HexAnima 1.2.0

Superseded by 1.2.2. The Snapdragon 8 Elite support described below turned out to be unreliable (garbage pictures after interrupted runs, noise at large sizes) and was withdrawn. Snapdragon 8 Gen 1 / 8+ Gen 1 phones are not affected.

HexAnima 1.2.0

New: experimental support for Snapdragon 8 Elite (Hexagon V79), for example Galaxy S25. Thanks to a tester who ran the builds on a Galaxy S25 and sent logs.

• The app now detects your NPU generation and downloads the matching text encoder / VAE files (1.2 GB the first time). Snapdragon 8 Gen 1 / 8+ Gen 1 phones keep working exactly as before.
• On the S25 the pictures come out correct. It is slower than expected for now: about 23 s for a 2.9B model at 512x768 (27 s with the first model load), against about 12 s on a Snapdragon 8+ Gen 1. Speed tuning for V79 is the next step.
• If you updated from 1.1.x on an 8 Elite phone and a model shows "Download" again, tap it: only the missing V79 files are downloaded, the model itself is not.

Also includes 1.1.1 and 1.1.2: the engine now starts on newer Android versions (the older C++ library from /vendor no longer breaks system tools or the helper programs).

Supported chips: Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475, tested); Snapdragon 8 Elite (SM8750, experimental). Snapdragon 8 Gen 2 and 8 Gen 3 are not supported yet.

Install: download HexAnima-1.2.0.apk, allow installs from unknown sources, open it. If you have an earlier version, install over it: models and history stay. Check with sha256sum -c SHA256SUMS.txt.
Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba or the model authors.

HexAnima 1.1.1

HexAnima 1.1.1

Fixed: the engine did not start on newer Android (for example Android 17). The log showed CANNOT LINK EXECUTABLE "setsid": cannot locate symbol _ZNSt3113hash_memory....

Cause: the start script set the library search path (/vendor/lib64) for every command it ran. Android's own tools (setsid, grep, sleep) then picked up an older C++ library from /vendor and failed to start on systems where the system library is newer. The path is now set only for HexAnima's own helper programs.

Nothing else changes. If you are on 1.1, just install this APK over it: your models and history stay.

Still true: only Snapdragon 8 Gen 1 / 8+ Gen 1 (SM8450 / SM8475, Hexagon V69) are supported. Newer chips (for example Snapdragon 8 Gen 2/3 and 8 Elite, SM8550 / SM8650 / SM8750) have a different NPU generation and the models will not run on them yet, even though the app can be started.

Install: download HexAnima-1.1.1.apk, allow installs from unknown sources, open it. Check with sha256sum -c SHA256SUMS.txt.
Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba or the model authors.

HexAnima 1.1

HexAnima 1.1

On-device Anima image generation for Snapdragon 8 Gen 1 / 8+ Gen 1 (Hexagon V69). No root needed. Models are downloaded in the app (Models).

• Rootless app (com.hexanima.npu), model manager with resumable verified downloads
• History, favourites, trash, batch generation, stop button, 13 languages, themes

Install: download HexAnima-1.1.apk, allow installs from unknown sources, open it. Verify with sha256sum -c SHA256SUMS.txt.
Free and non-commercial. Contains Qualcomm AI Engine Direct runtime libraries under the Qualcomm AI Stack License (see Info -> Licenses).
Unofficial: not affiliated with CircleStone Labs, NVIDIA, Qualcomm, Alibaba or the model authors.