This website requires JavaScript.
1bd2745ad0
0.4.6: Load Video uint8 + max_side (anti-OOM 4K) + mensaje OOM correcto; cap default 33
neuregex
2026-06-13 17:39:42 -06:00
7c9afcdb32
0.4.5: salida de video BerniniRSaveVideo (mp4/webm/gif/webp) + preview; reemplaza SaveAnimatedWEBP
neuregex
2026-06-13 17:18:01 -06:00
b8ae9781c7
0.4.4: patch_embedding tolerante a dtype - arregla t2v/rv2v incondicional en 1.3B bf16
neuregex
2026-06-13 16:58:11 -06:00
ee6a7c232a
0.4.3: preview de video + boton de carga (mp4/webp/gif) en Load Video
neuregex
2026-06-13 06:06:18 -06:00
894f0ac936
0.4.2: Load Video soporta mp4/mov/avi/mkv/webm (PyAV/OpenCV/imageio) + av dep
neuregex
2026-06-13 05:57:02 -06:00
c281fc923a
0.4.1: evita falsos positivos del scan (reword 'su' en string + quita os.environ)
neuregex
2026-06-13 04:49:39 -06:00
a83961fe2b
release: v0.4.0 - backend nativo (GGUF dual-expert + 1.3B), TE GGUF, resolución acotada
neuregex
2026-06-11 04:34:36 -06:00
cbd6f7f6dd
workflows: elimina diffusers/API legacy + tests nativos (quedan los 8 grafos v0.4)
neuregex
2026-06-10 02:07:57 -06:00
52d5393b6e
workflows: GGUF text encoder por defecto + cap de resolución (ImageScale)
neuregex
2026-06-10 02:07:48 -06:00
45bc96d0b4
fix: patch_embedding input al dtype de computo (no float32) - arregla bf16 puro (1.3B); no-op fp8/GGUF
neuregex
2026-06-10 01:10:58 -06:00
8e0dee965b
feat(workflows): set completo GGUF/1.3B - t2v 14B + t2v/v2v/rv2v 1.3B
neuregex
2026-06-10 01:03:19 -06:00
d832069445
docs: README - seccion Models-por-VRAM (14B GGUF + 1.3B) y edicion dual-expert (i2i/v2v/rv2v) validada
neuregex
2026-06-10 00:49:17 -06:00
72fa8d436c
feat: workflow i2i Bernini-R 1.3B (monolitico single-expert, sin model_low)
neuregex
2026-06-09 23:59:59 -06:00
79fedf7dc3
feat: workflow rv2v (video+referencia dual-expert GGUF) + auto-cleanup del experto low (evita thrashing VRAM entre workflows)
neuregex
2026-06-06 23:57:52 -06:00
ac363feb6c
fix(streams): normaliza el latente fuente con process_latent_in (mean/std Wan); el stream iba en espacio crudo y el target no seguia la fuente
neuregex
2026-06-06 22:22:44 -06:00
9223f6d8b6
feat(M4/v2v): ImageScale 768x576 (div16) + mode v2v_apg en el workflow de video
neuregex
2026-06-06 21:23:56 -06:00
ead504a363
fix(M4/v2v): freqs desde el grid de parches (2*Hg,2*Wg) - arregla mismatch tokens/freqs con dims impares de video; no-op i2i/t2v
neuregex
2026-06-06 20:28:32 -06:00
9d3cf99066
ux: BerniniRLoadVideo -> selector de archivo + boton de subida (image_upload), como LoadImage
neuregex
2026-06-06 20:15:59 -06:00
6c05e327a8
feat(M4/v2v): workflow GGUF dual-expert para edicion de video (LoadVideo->VAEEncode->stream)
neuregex
2026-06-06 20:11:09 -06:00
5c8cdc7d56
fix(M4): fija current_patcher del experto low (pre_run) para calc_cond_batch en el switch
neuregex
2026-06-06 19:56:52 -06:00
8fdb000e66
docs(M4): README -> link Bernini-R-GGUF + Discord + camino GGUF dual-expert; feat: switch dual-expert + workflow GGUF
neuregex
2026-06-06 06:09:37 -06:00
cbd6f9fb9b
feat(M4): switch dual-expert por sigma en el guider + nodo ApplyPatches + workflow GGUF dual
neuregex
2026-06-06 05:58:41 -06:00
6a0f675805
feat(M4): fp8->bf16 streaming upcast + BerniniRApplyPatches (camino GGUF)
neuregex
2026-06-06 04:10:26 -06:00
d6ed16a2c3
feat(native M3b): BerniniRGuider (APG + CFG encadenada, 7 modos) sobre CFGGuider; workflow i2i con SamplerCustomAdvanced
neuregex
2026-06-06 03:33:22 -06:00
5b01b760b9
fix(native M3a): concatenar freqs por el eje de tokens (dim 1), no por el broadcast de cabezas (dim 2)
neuregex
2026-06-06 03:10:27 -06:00
4c5d134e75
fix(native M3a): expandir stream fuente al batch del target (CFG) en el concat
neuregex
2026-06-06 02:54:12 -06:00
d988116080
feat(native M3a): nodo BerniniRSourceStream + forward device-fix + workflow i2i (activa el stream-concat de M2)
neuregex
2026-06-06 02:49:06 -06:00
2582f5cc9c
feat(native M2): patches Bernini via add_object_patch - src-id RoPE real (validado) + stream-concat (no-op en t2v)
neuregex
2026-06-06 02:07:50 -06:00
4270abca73
fix(native): set_inference_dtype manual_cast=bf16 -> ops castean fp8/bf16 al vuelo (v0.4 M1)
neuregex
2026-06-06 00:53:17 -06:00
e7c2563886
feat(native): loader meta+assign (esqueleto 0-RAM, evita torch.empty fp8) (v0.4 M1)
neuregex
2026-06-06 00:36:51 -06:00
eb80291e67
feat(native): loader MODEL con lectura fp8-safe (esquiva crash load_torch_file torch2.8/win) (v0.4 M1)
neuregex
2026-06-05 23:55:53 -06:00
a56bc058dc
fix(offload): aparcar experto inactivo en disco (mmap) en vez de copiar 14.6GB a RAM (v0.3.14)
neuregex
2026-06-05 17:55:01 -06:00
171453bc42
fix(model): to_idle no mueve si ya esta en CPU (evita access violation sobre fp8 mmap read-only) (v0.3.13)
neuregex
2026-06-05 17:24:48 -06:00
73770e1278
fix(loader): mmap solo-lectura (sin reservar commit) -> arregla os error 1455 al cargar UMT5 (v0.3.12)
neuregex
2026-06-05 17:12:29 -06:00
026ca26793
fix(loader): detectar fp8 por header (robusto) -> 2o experto usa mmap, no el f.read bf16 (v0.3.11)
neuregex
2026-06-05 17:05:16 -06:00
178deee85b
fix(loader): mmap perezoso para fp8 (no clona) -> sin MemoryError ni segfault (v0.3.10)
neuregex
2026-06-05 17:00:29 -06:00
2b2dc5ef12
fix(loader): cargar fp8 a mano (bytes->view) esquiva access violation de safetensors fp8 en torch2.8/win; barra xet limpia (v0.3.9)
neuregex
2026-06-05 16:50:09 -06:00
648ba23020
ux: progreso xet limpio (solo cambios + latido) v0.3.8
neuregex
2026-06-05 16:37:32 -06:00
5fa835dd89
perf: no desactivar xet (era el cuello de botella); motor xet auto si hf_xet (v0.3.7)
neuregex
2026-06-05 15:54:56 -06:00
31a548e3cb
feat: barra de progreso real por bytes (vigila disco) + guard de integridad + xet-off interno (v0.3.6)
neuregex
2026-06-05 15:35:40 -06:00
04e0e557e7
fix: progreso por bytes + verificacion de integridad + xet-off interno (v0.3.5)
neuregex
2026-06-05 15:20:46 -06:00
8e2f087bc0
feat(loader): real byte-progress download (%, MB/s, ETA) + disable xet by default
neuregex
2026-06-05 19:01:18 +00:00
5ebccb615c
chore: bump 0.3.3 (fix widget-order blocks_to_swap + ui/*.json regenerados)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 12:12:57 -06:00
d46c7f2c8b
chore(workflows): regenera ui/*.json contra el nodo actual (incluye blocks_to_swap)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 12:12:42 -06:00
44761a960d
fix(loader): blocks_to_swap al final del bloque optional (orden de widgets)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 12:09:00 -06:00
42ba6363f2
chore(repo): elimina el harness de Modal y toda referencia — solo nodo funcional
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 02:42:40 -06:00
ec8deb8d3c
chore: bump 0.3.1 (paquete publicado sin el harness de validación)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 02:25:21 -06:00
31e4587dd4
chore(registry): .comfyignore — excluye modal/, tests/, .github/ del paquete
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 02:25:20 -06:00
b76a9feb02
chore: bump 0.3.0 (block-swap, lower VRAM to ~12GB)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 01:57:55 -06:00
c4b7ec0c6f
docs(block-swap): VRAM/velocidad medidos + guía blocks_to_swap por GPU
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 01:18:06 -06:00
4355e1fb22
feat(modal): bswap_bench — pico VRAM + velocidad por block_swap (i2i/t2v)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 01:17:18 -06:00
9817fbbfe0
test(block-swap): gate de correctitud VERDE — i2i bs=0 vs bs=30 bit-idéntico
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 00:54:21 -06:00
deb75c9a75
fix(block-swap): synchronous block transfers (non_blocking caused GPU memory-reuse corruption)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-05 00:03:34 -06:00
c9d27f2152
feat(block-swap): wire blocks_to_swap (sampler _activate, ModelLoader widget, run harness)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 23:47:48 -06:00
e808e2525a
feat(block-swap): BerniniExpert block_swap + to_active/to_idle + _run_blocks streaming
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 23:45:10 -06:00
ce98ba2715
fix(fp8): restrict HF stale-delete to transformers + clear upload cache (restore text_encoder)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 20:34:52 -06:00
52e72a8757
fix(fp8): clean dst before re-save + delete stale HF shards on upload (one-file bundle)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 20:28:57 -06:00
d62abaabc5
feat: one-file fp8 shards (max_shard_size=30GB) + run --input-image/--omega-* overrides
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 20:21:07 -06:00
41e9579be5
feat(workflows): UI-format (drag-drop) for t2v/i2i, generated + GUI-validated in real ComfyUI; README API vs UI note
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 19:58:20 -06:00
27a5d1f378
feat(modal): node_info + gui_smoke (Playwright) — validate node registration, UI graph load/BR_PATH, i2i Run; _Comfy --cpu
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 19:58:19 -06:00
388fdbad17
chore: bump to 0.2.0 (fp8 auto-download bundle, single-source BR_PATH, built-in video loader)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 19:37:10 -06:00
61ace95032
fix(graph): single weights source via BR_PATH output (VAE/TextEncode wire to Load Model); harness hf_source/download_dir
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 19:10:52 -06:00
e9bb6452e4
docs(weights): point README at fp8 HF repo + source/auto_download/download_dir; built-in video loader note
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:48:53 -06:00
0381b527ae
fix(modal): upload_fp8 uses existing hf-secret + robust token var lookup
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:43:15 -06:00
a5c3da7eb8
feat(modal): upload_fp8 + fp8 model card (HF upload from Volume)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:37:28 -06:00
8d42791ff2
feat(modal): build_fp8 — build self-contained fp8 bundle on the Volume (no GPU)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:16:14 -06:00
ea58b06acd
feat(loader): source/auto_download/download_dir widgets + HF auto-download; wire workflows + Modal harness (model_subdir)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:14:51 -06:00
15a958c9ed
feat(fp8): load_experts auto-detects pre-fp8 repo and preserves e4m3 (init_empty_weights + assign)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:10:04 -06:00
32786382c0
feat(fp8): tools/save_fp8_diffusers.py — build self-contained fp8 bundle + verify e4m3 on disk
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:08:15 -06:00
fee5ffee8c
refactor(fp8): split quantize_fp8_ into cast_fp8_ + apply_fp8_forward_ (preserve pre-fp8 weights)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 18:06:48 -06:00
c56464f1a7
docs: rewrite README in English; update validation status + VRAM with real Modal results
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 16:49:17 -06:00
44738dfaf4
chore: prepare ComfyUI Registry publish (LICENSE, publish workflow, README badges)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 16:34:35 -06:00
e781b2913f
feat(modal): gpu parametrizable + reporte VRAM; valida i2i fp8 en A10 24GB (16.7GB pico)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 15:48:02 -06:00
08110f4529
feat(modal): run --gen-input genera paisaje de prueba para validar i2i (src-id cross-stream)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 14:30:35 -06:00
6eec806259
fix: bf16 dtype IO + pins Modal (transformers 4.49, torchaudio cu124) + A100-80GB
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 10:02:07 -06:00
e21c1d3084
test: ancla de fidelidad rope src_id=0 ≡ Wan2.2 stock (C1)
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 08:52:02 -06:00
1e3f771ca7
test: wiring CPU end-to-end + fix pytest import-mode para custom node
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 08:39:17 -06:00
1dfe723e1a
scaffold
Sammir Alejandro Femat Robles and Claude Opus 4.8
2026-06-04 08:17:56 -06:00
neuregex