Spaces:

Luminia
/

Anima-2B-CPU

Running

App Files Files Community

Anima-2B-CPU

Commit History

May 19

07:20
fix: use anime model (Anima-V1-Turbo) not semi-realistic (nArnima) Running

Nekochu

707110b
07:19
exact copy of working private Space: Docker + n-Arno V1 Turbo + CFG 1 (899s proven)

Nekochu

c49a405
06:28
Gradio SDK + n-Arno V1 Turbo AIO + CFG 1 + pre-built binary

Nekochu

d4579d5

May 18

23:39
remove 832x1216 resolution

Nekochu

7fe9427
14:24
restore cache-mode spectrum + fa + offload (18GB RAM, only 3GB used)

Nekochu

feda154
12:38
test: no BLAS at all (old 854s Docker had BLAS silently disabled)

Nekochu

da2881d
08:31
fix: compile with -j1 to avoid OOM during Docker build

Nekochu

9dfcabd
07:38
fix: add pkg-config for BLAS detection

Nekochu

7eb2266
07:33
fix: remove max_size=1, was rejecting queued requests

Nekochu

b8327a3
07:31
back to Docker: v1.0 model, OpenBLAS BLAS, CFG 2 er_sde, CLI+API+MCP

Nekochu

9f3da1a
04:14
cache compiled sd-cli with OpenBLAS

Nekochu

Verified signature
d6719f5
04:03
auto-cache sd-cli binary in vendor/ after first compile

Nekochu

fb913db
03:50
try OPENBLAS_NUM_THREADS=2 for speed (Codex fixes handle RAM)

Nekochu

2a51dde
03:13
Codex fixes: disk-backed logs, malloc_trim, lazy imports, no _on_unload

Nekochu

642889e
03:03
fix OOM: disable SSR, limit OpenBLAS to 1 thread

Nekochu

250cdf5
02:51
fix: only kill sd-cli for browser disconnects, not API/MCP

Nekochu

9288693
02:09
fix: distinguish user disconnect from real OOM (both exit -9)

Nekochu

91aa06f
01:45
fix OOM: drop cache-mode spectrum + offload-to-cpu to reduce RAM

Nekochu

d1362f7
01:32
fix: use GGML_BLAS=ON + GGML_BLAS_VENDOR=OpenBLAS (new ggml API)

Nekochu

3d33d2f
01:15
compile sd-cli from source with OpenBLAS, add elapsed to errors

Nekochu

2534921

May 17

13:15
fix: restore cache-mode spectrum (none not valid)

Nekochu

72664f6
13:10
fix OOM: drop --fa, cache-mode none to reduce RAM

Nekochu

ce35f12
12:16
fix: extract libstable-diffusion.so, set LD_LIBRARY_PATH, fix api_name

Nekochu

f2d38f5
12:10
v1.0: ditch Docker, Gradio SDK, CLI+API+MCP, CFG 2 er_sde 8 steps

Nekochu

c9c8db3

May 15

10:43
Include resolution/steps/seed in OOM error message

Nekochu

db73a8e
10:36
Update repo name to Anima-v1.0-Turbo-AIO-Q4_K

Nekochu

100220e
10:14
Update to Anima Base v1.0 + Turbo AIO

Nekochu

ea4db51

May 9

20:26
Add gr.Request to unload + gc.collect after kill

Nekochu

c49857d
20:21
Drop 832x1216 (OOM on free CPU), add queue(concurrency=1) for MCP

Nekochu

32b078b
19:20
Add MCP server support, type hints, and docstrings for generate()

Nekochu

509b9a7
19:17
Remove redundant size from header

Nekochu

d0b4d4d
19:16
Default 8 steps (Turbo LoRA spec), slider 8-16, link source models

Nekochu

d6d1e0f
18:58
Switch to n-Arno AIO Q4_K GGUF (single 1.79 GB file, restores GGUF speed)

Nekochu

6da7294
18:38
Switch to Anima Preview3 + new Turbo LoRA (fixes oversaturation reported in #1)

Nekochu

d528a08

Apr 30

18:03
lock resolution max 768x512, 768x768 OOMs on free tier

Nekochu

b357b62
17:24
try dpm++2m sampler (lower RAM than er_sde)

Nekochu

3959a08
17:22
OOM error message for code -9, suggest 512x512

Nekochu

8a4f075
16:30
timeout 3h, remove est warning

Nekochu

4ff9ccb
15:45
fix warning wording

Nekochu

172ebc4
15:41
fix sampler: euler -> er_sde (recommended for turbo LoRA)

Nekochu

917080b
15:33
fix seed crash, lock steps/cfg, kill subprocess on disconnect, concurrency limit, time warnings

Nekochu

767b311
15:24
fix seed crash, cap resolution 768x768, steps max 12, concurrency limit 1, time estimate

Nekochu

b9ec9f4

Mar 27

04:52
Edit readme color

Nekochu

bb1c05e

Mar 23

15:12
add censored to neg prompt

Nekochu

34f28f8
14:52
add negative prompt, show timing estimate

Nekochu

ec4bc1d
11:30
fix: server_port not port

Nekochu

de16836
08:54
fix: clone with submodules

Nekochu

56d5f4e
07:39
switch to Docker with compiled sd.cpp

Nekochu

e1ae826

Mar 22

07:09
fix: use git sd.cpp for Anima support, fix theme

Nekochu

8852c51

Mar 21

05:48
Initial Anima 2B CPU Space with Turbo LoRA

Nekochu

1e35cb0

Commit History

May 19

fix: use anime model (Anima-V1-Turbo) not semi-realistic (nArnima) Running

exact copy of working private Space: Docker + n-Arno V1 Turbo + CFG 1 (899s proven)

Gradio SDK + n-Arno V1 Turbo AIO + CFG 1 + pre-built binary

May 18

remove 832x1216 resolution

restore cache-mode spectrum + fa + offload (18GB RAM, only 3GB used)

test: no BLAS at all (old 854s Docker had BLAS silently disabled)

fix: compile with -j1 to avoid OOM during Docker build

fix: add pkg-config for BLAS detection

fix: remove max_size=1, was rejecting queued requests

back to Docker: v1.0 model, OpenBLAS BLAS, CFG 2 er_sde, CLI+API+MCP

cache compiled sd-cli with OpenBLAS

auto-cache sd-cli binary in vendor/ after first compile

try OPENBLAS_NUM_THREADS=2 for speed (Codex fixes handle RAM)

Codex fixes: disk-backed logs, malloc_trim, lazy imports, no _on_unload

fix OOM: disable SSR, limit OpenBLAS to 1 thread

fix: only kill sd-cli for browser disconnects, not API/MCP

fix: distinguish user disconnect from real OOM (both exit -9)

fix OOM: drop cache-mode spectrum + offload-to-cpu to reduce RAM

fix: use GGML_BLAS=ON + GGML_BLAS_VENDOR=OpenBLAS (new ggml API)

compile sd-cli from source with OpenBLAS, add elapsed to errors

May 17

fix: restore cache-mode spectrum (none not valid)

fix OOM: drop --fa, cache-mode none to reduce RAM

fix: extract libstable-diffusion.so, set LD_LIBRARY_PATH, fix api_name

v1.0: ditch Docker, Gradio SDK, CLI+API+MCP, CFG 2 er_sde 8 steps

May 15

Include resolution/steps/seed in OOM error message

Update repo name to Anima-v1.0-Turbo-AIO-Q4_K

Update to Anima Base v1.0 + Turbo AIO

May 9

Add gr.Request to unload + gc.collect after kill

Drop 832x1216 (OOM on free CPU), add queue(concurrency=1) for MCP

Add MCP server support, type hints, and docstrings for generate()

Remove redundant size from header

Default 8 steps (Turbo LoRA spec), slider 8-16, link source models

Switch to n-Arno AIO Q4_K GGUF (single 1.79 GB file, restores GGUF speed)

Switch to Anima Preview3 + new Turbo LoRA (fixes oversaturation reported in #1)

Apr 30

lock resolution max 768x512, 768x768 OOMs on free tier

try dpm++2m sampler (lower RAM than er_sde)

OOM error message for code -9, suggest 512x512

timeout 3h, remove est warning

fix warning wording

fix sampler: euler -> er_sde (recommended for turbo LoRA)

fix seed crash, lock steps/cfg, kill subprocess on disconnect, concurrency limit, time warnings

fix seed crash, cap resolution 768x768, steps max 12, concurrency limit 1, time estimate

Mar 27

Edit readme color

Mar 23

add censored to neg prompt

add negative prompt, show timing estimate

fix: server_port not port

fix: clone with submodules

switch to Docker with compiled sd.cpp

Mar 22

fix: use git sd.cpp for Anima support, fix theme

Mar 21

Initial Anima 2B CPU Space with Turbo LoRA