전체 검색

AI 모델 2119

• SAME-L↗

  Stability AI

  Latent representations are at the heart of the majority of modern generative models. In the audio domain they are typically produced by a neural-audio-codec autoencoder. In this work we introduce SAME (Semantically Aligned Music autoEncoder), a transformer-based autoencoder for stereo music and general audio that reaches a 4096x temporal compression ratio (roughly twice the current standard) while maintaining excellent reconstruction quality and strong downstream generative performance. We achieve this by combining a set of semantic regularisation approaches with phase-aware reconstruction losses. The architecture also delivers substantial computational cost benefits, through both its high compression ratio and its reliance on well-optimised transformer primitives. Two variants (a large SAME-L and a CPU-deployable SAME-S) are released in open-weights form. This model can be used with: 1. the `stable-audio-3` inference and fine-tuning library 2. the `stable-audio-tools` research library Our dataset consists of ~19,500 hours of licensed production audio from AudioSparx which includes a 66/25/9% mix of music, sound effects, and instrument stems.

• occamy-1.0-APEX-GGUF↗

  Accio-Lab

  Occamy-1.0 · APEX GGUF Three plain profiles · four calibrated profiles, including Mini Collection · Checkpoint explorer · Project · Paper Occamy-1.0 quantized with the LocalAI team's APEX profiles. APEX allocates GGUF precision by tensor role and layer position: routed experts, shared experts and attention receive different allocations. The three plain profiles and four calibrated `I-` profiles use the pinned upstream 40-layer configurations. Every file is generated directly from the same BF16 source. The calibrated group adds I-Compact, I-Quality, I-Balanced and I-Mini. Profile names describe allocation strategies; file size does not follow the name order. These are weight-file sizes, not peak memory requirements. Allow additional memory for the runtime, context and image input.

• AnyPose↗

  lilylilith

  Creating poses for Qwen Image Edit (2511) or its variants can be difficult. Even with openpose being baked into the model, it can still look like something is off with the result, such as the depth being incorrect or the pose not fully matching the character, with distortions that you wouldn't expect. Additionally, creating them by hand can take a bit too long, especially if you are going the blender openpose rig route. What if instead you could just skip openpose entirely and write a simple (but albeit lengthy) prompt to copy the pose of any image on your computer? Well, this is what the AnyPose LoRAs attempt to do. Made in mind with the new Qwen Image Edit 2511 lightning LoRA for fast inference, with just a single reference image as a pose guide, you can pilot any image to follow that pose. No control net needed. Use 4-step lightning LoRA for Qwen Image Edit 2511. Set both of the AnyPose LoRA weights (base & helper) strength to 0.7. Upload two images, one with the initial input character, and another with the pose you want copied to the first character. Use the following prompt, "Make the person in image 1 do the exact same pose of the person in image 2. Changing the style and background of the image of the person in image 1 is undesirable, so don't do it. The new pose should be pixel accurate to the pose we are trying to copy. The position of the arms and head and legs should be the same as the pose we are trying to copy. Change the field of view and angle to match exactly image 2. Head tilt and eye gaze pose should match the person in image 2." If additional context is needed (such as clothing that is unseen from the model), prompt it in at the end. Background not staying the same as the initial input image? Append the following to the end "Remove the background of image 2, and replace it with the background of image 1."

• samantha-mistral-7b-i1-GGUF↗

  mradermacher

• mt5-large↗

  Google

  mT5 is pretrained on the mC4 corpus, covering 101 languages: Afrikaans, Albanian, Amharic, Arabic, Armenian, Azerbaijani, Basque, Belarusian, Bengali, Bulgarian, Burmese, Catalan, Cebuano, Chichewa, Chinese, Corsican, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Haitian Creole, Hausa, Hawaiian, Hebrew, Hindi, Hmong, Hungarian, Icelandic, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish, Kyrgyz, Lao, Latin, Latvian, Lithuanian, Luxembourgish, Macedonian, Malagasy, Malay, Malayalam, Maltese, Maori, Marathi, Mongolian, Nepali, Norwegian, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Samoan, Scottish Gaelic, Serbian, Shona, Sindhi, Sinhala, Slovak, Slovenian, Somali, Sotho, Spanish, Sundanese, Swahili, Swedish, Tajik, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, West Frisian, Xhosa, Yiddish, Yoruba, Zulu. **Note**: mT5 was only pre-trained on mC4 excluding any supervised training. Therefore, this model has to be fine-tuned before it is useable on a downstream task.

• Index-Echo-S2TT-GGUF↗

  audio-cpp

  GGUF packages for native speech-to-text translation with audio.cpp. Chinese speech produces timestamped source transcripts and English, Spanish, or Japanese translations. This port supports **S2TT only**; it does not generate translated speech. **There is room to improve 9B output parity and quantized quality**. Focused PRs with reproducible C++/Python comparisons are welcome. Note that 2B and 9B use the same C++ translation path, including the prompt, audio encoder, decoder implementation, and greedy sampler. 2B orig is parity safe, while 9B orig has drifts. 2B has 24 layers and tied token/output embeddings; 9B has 32 layers and a separate LM head. That can change how sensitive their outputs are to numerical differences, but I have not proved that it explains the 9B mismatch. These files were converted from the official 2B checkpoint at `72fd8bdc9bf8251f1c35f6d6289a950eda154395` and 9B checkpoint at `05f86cb7a38684916e193282b6cc02462ace4043`. Each GGUF embeds the audio tower, connector, text decoder, tokenizer, configuration, and audio.cpp model spec. The unused vision weights are not included.

• wide_resnet50_2.tv2_in1k↗

  timm

  This model features: * ReLU activations * single layer 7x7 convolution with pooling * 1x1 convolution shortcut downsample

• qwen2.5-7b-instruct-arabic-yt-merged↗

  HarithSami

• eomt-dinov3-ade-semantic-large-512↗

  tue-mps

  This is the **large** variant of the EoMT-DINOv3 model trained for **semantic segmentation** on ADE20K at **512×512** resolution. **EoMT (Encoder-only Mask Transformer)** is a Vision Transformer (ViT) architecture designed for high-quality and efficient image segmentation. It was introduced in the CVPR 2025 highlight paper: **Your ViT is Secretly an Image Segmentation Model** > **Key Insight**: Given sufficient scale and pretraining, a plain ViT along with a few additional parameters can perform segmentation without the need for task-specific decoders or pixel fusion modules. The same model backbone supports semantic, instance, and panoptic segmentation with different post-processing.

• sd-vae-ft-mse-flax↗

  enterprise-explorers

  These weights are intended to be used with the 🧨 diffusers library. If you are looking for the model to use with the original CompVis Stable Diffusion codebase, come here. We publish two kl-f8 autoencoder versions, finetuned from the original kl-f8 autoencoder on a 1:1 ratio of LAION-Aesthetics and LAION-Humans, an unreleased subset containing only SFW images of humans. The intent was to fine-tune on the Stable Diffusion training set (the autoencoder was originally trained on OpenImages) but also enrich the dataset with images of humans to improve the reconstruction of faces. The first, _ft-EMA_, was resumed from the original checkpoint, trained for 313198 steps and uses EMA weights. It uses the same loss configuration as the original checkpoint (L1 + LPIPS). The second, _ft-MSE_, was resumed from _ft-EMA_ and uses EMA weights and was trained for another 280k steps using a different loss, with more emphasis on MSE reconstruction (MSE + 0.1 * LPIPS). It produces somewhat ``smoother'' outputs. The batch size for both versions was 192 (16 A100s, batch size 12 per GPU). To keep compatibility with existing models, only the decoder part was finetuned; the checkpoints can be used as a drop-in replacement for the existing autoencoder.