Open-Weight Models Under Apache License 2.0

PP-OCRv6: From 1.5M to 34.5M Parameters, Surpassing Billion-Scale VLMs on OCR Tasks PP-OCRv6 is a lightweight OCR system that combines architectural innovation with data-centric optimization. It redesigns the backbone, detection neck, and recognition neck around a unified MetaFormer-style building block with structural reparameterization. Three model tiers (medium, small, tiny) share the same block primitives, covering deployment scenarios from server to edge. 1. Unified and Scalable Model Family: A three-tier OCR model family spanning 1.5M to 34.5M parameters. PP-OCRv6medium achieves 86.2% detection Hmean and 83.2% recognition accuracy, outperforming PP-OCRv5server by +4.6% and +5.1%…

Model · Text classification

pplx-decider-v1-27b-EXL3

ramGPT

EXL3 4.00 bpw conversion of perplexity-ai/pplx-decider-v1-27b. pplx-decider-v1-27b does not use the normal LM-head token-generation path for its final answer. It produces a hidden state and applies the model's custom readout.safetensors decision head to obtain option probabilities. The repository therefore includes pplxdeciderexl3.py, which registers the model architecture and performs the custom decision readout. This model is not recommended for standard TabbyAPI usage. A local smoke test with TabbyAPI and ExLlamaV3 1.5.2 failed during startup with: More importantly, simply registering the architecture is not sufficient for normal /v1/chat/completions behavior: TabbyAPI's standard…

This model is a fine-tuned version of qing-yao/ppt-pythia-160m-uniform250-previousmse-seed324-stage1 on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.001 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosinewithminlr - lrschedulerwarmupsteps: 500 - trainingsteps: 10000 - Transformers 5.4.0 - Pytorch 2.8.0+cu128 - Datasets 3.2.0 - Tokenizers 0.22.1

This model is a fine-tuned version of qing-yao/ppt-pythia-160m-uniform250-previousmse-seed324-stage1 on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.001 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosinewithminlr - lrschedulerwarmupsteps: 500 - trainingsteps: 10000 - Transformers 5.4.0 - Pytorch 2.8.0+cu128 - Datasets 3.2.0 - Tokenizers 0.22.1

This model is a fine-tuned version of EleutherAI/pythia-160m on an unknown dataset. The following hyperparameters were used during training: - learningrate: 0.001 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosinewithminlr - lrschedulerwarmupsteps: 13 - trainingsteps: 250 - Transformers 5.4.0 - Pytorch 2.8.0+cu128 - Datasets 3.2.0 - Tokenizers 0.22.1

This model is a fine-tuned version of EleutherAI/pythia-160m on an unknown dataset. The following hyperparameters were used during training: - learningrate: 0.001 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosinewithminlr - lrschedulerwarmupsteps: 13 - trainingsteps: 250 - Transformers 5.4.0 - Pytorch 2.8.0+cu128 - Datasets 3.2.0 - Tokenizers 0.22.1

This model is a fine-tuned version of EleutherAI/pythia-160m on an unknown dataset. The following hyperparameters were used during training: - learningrate: 0.001 - trainbatchsize: 16 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 32 - lrschedulertype: cosinewithminlr - lrschedulerwarmupsteps: 13 - trainingsteps: 250 - Transformers 5.4.0 - Pytorch 2.8.0+cu128 - Datasets 3.2.0 - Tokenizers 0.22.1

static quants of https://huggingface.co/PreciseCoder/PreciseCoder-9B For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality…

Model

preemption-safe-training-template

Viswaretas kotra

A minimal PyTorch training loop that survives spot / interruptible GPU preemption. Copy train.py and swap in your model and data. 1. Keeps all recovery state (weights, optimizer, RNG, step) in one directory: $NODUSSTATEDIR or./state. 2. Writes checkpoints atomically (temp file, then os.replace). 3. Resumes from the last checkpoint on startup. 4. Saves on a timer, and again immediately when the platform sends a checkpoint request. Nodus copies the state directory off the machine before a reclaim and restores it on the replacement, so preemptions cost minutes instead of the whole run. New accounts get a $30 starter grant (valid for 30 days), so the first runs need no card. Add --dry-run to…

每个目录是一次训练：a.pt 是可训练状态（仅 FFN 三矩阵，264.2M），a.step1000.pt 是中间存点， 门控的 run 另有 folded.pt（折叠成逐层宽度，评测用这个），train.log 含每步损失与留出集曲线。 它是旧批次的门控 75%（原名 c32kgate2304）：旧数据池（150M token，同一份数据过 3.49 遍， 预期分数与逐任务明细在仓库的 EVALUATING.md，摘要： 评测它必须用 --window 128，因为它是按 128–256 训的。用 512 评它会低估 —— 那是它没训练过的 regime。 不要把它的分数当作方法的结果。 新批次的门控 75%（b2gate2304）用新数据池和连续窗口重训，

Model · Text generation

prism-roleplay-1.5-small-fast

Vertex AGI

The Prism Roleplay 1.5 Small recipe on PrismML's 1-bit Bonsai-8B: a standalone model, faster and smaller than the 4-bit original, with a small quality gap A standalone model, not an adapter: load it directly with mlx-lm. It is Prism Roleplay 1.5 Small's training recipe (same data, hyperparameters and system prompt) applied to prism-ml/Bonsai-8B-mlx-1bit, a 1-bit (g128) build of Qwen3-8B. How the LoRA was merged into a 1-bit model. A plain merge re-quantizes every layer to 1 bit, and the small LoRA update disappears in the rounding (we measured it: merged held-out loss 3.415 = untuned base 3.405). So the 16 layers the LoRA touched (all their attention and MLP projections) are merged at fp16…

Model · Text generation

PrismaCoder-3B

Sandeep Thummala

PrismaCoder-3B is a Qwen2.5-Coder-3B-Instruct fine-tune specialized for generating and editing Prisma ORM schemas and client code from natural-language instructions. It was trained with a full SFT → DPO → GRPO pipeline, using a verifier (gate) as both the data-quality filter and the reinforcement-learning reward signal. stage, LoRA weights merged into the base model between stages iterative self-play RL round (attempted; see Limitations) weights for release All stages used seed 20260830 and ran on a single Kaggle T4 GPU (16GB) with gradient checkpointing and paged 8-bit AdamW. verifier (LightValidator, "V0" in the project's internal gate numbering): it checks that generated Prisma…

Model · Text to image

prismatic-dark-fantasy

Lucas Lagos

These are Llagos2025/prismatic-dark-fantasy DreamBooth LoRA weights, trained on krea/Krea-2-Raw. The weights were trained using DreamBooth with the Krea 2 diffusers trainer. Krea 2 ships as two checkpoints: RAW (the non-distilled base you fine-tune on) and Turbo (an 8-step distilled checkpoint for fast, high-quality inference). Train your LoRA on RAW and run it on Turbo — LoRAs trained on RAW express strongly on Turbo. You should use TOK to trigger the image generation. Download the.safetensors LoRA in the Files & versions tab. For more details, including weighting, merging and fusing LoRAs, check the documentation on loading LoRAs in diffusers [TODO: provide examples of latent issues and…

Model · Image to video

ProAR

Shen Linghui

ProAR solves reasoning tasks by autoregressively generating a sequence of visual states. It equips autoregressive video models with two complementary components: - Outcome Guidance predicts a future goal frame to guide current-chunk generation. - Transition Guidance aligns the current representation with the clean next-chunk representation during training. ProAR is built upon Wan2.2-TI2V-5B. hf download LinghuiShen/ProAR \ --local-dir checkpoints/ProAR

Model

Prochlo-190M-SFT

CyanoAI

Supervised fine-tuned version of Prochlo-190M-Base — a 190.5M-parameter dense English SLM trained from scratch. All variants share architecture, tokenizer, and vocabulary. 26-item behavioral rubric (format adherence, anti-sycophancy, code, concision): 61.5% pass. Strong on code instructions and factual QA. Weak on concise replies, long multi-turn coherence, and arithmetic. English only — Chinese input produces garbage by design. The chat format is plain text with control tokens, not a chat template: Note: weights are sharded across 32 safetensors files with a model.safetensors.index.json; frompretrained handles this transparently. Apache-2.0

Model

project-multitask

Robert Jones

Working implementation of Beit for Multitask using a huge configuration. The repository focuses on transparent code and repeatable smoke tests; benchmark claims are deliberately omitted. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses adafactor with a cosine schedule. These are starting values in the script, not evidence of a…

weighted/imatrix quants of https://huggingface.co/Naphula/Promethean-Dawn-26B-A4B For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/Promethean-Dawn-26B-A4B-GGUF This is a vision model - mmproj files (if any) will be in the static repository. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on…

A fast, multilingual classifier that flags prompt injection and jailbreak attempts, both in user messages (direct) and in untrusted content an AI agent reads: emails, web pages, documents, RAG chunks, and tool/API outputs (indirect). their clean counterparts, so it looks for instructions aimed at the AI, not for scary words. - Low false-alarm rate on look-alike benign text: 92.9% on NotInject, 99.2% on OR-Bench-hard. half the size and the same decisions as fp32 on our checks), transformers.js. Labels: SAFE (0) and INJECTION (1). This is the same convention as protectai/deberta-v3-base-prompt-injection-v2, so the model is a drop-in replacement in code and tools built for that one. INJECTION…

A compact multilingual detector for prompt injection and jailbreak attempts (large language model guardrails). Given any user prompt, tool output, or document excerpt, it outputs the probability that the text is an attack on the LLM's instructions. Existing public injection detectors are either English-only (e.g. protectai's deberta models) or released under licenses many organizations cannot use (Meta's PromptGuard line). This model is permissively licensed, works in 17 languages, ships with its multilingual training corpus and is evaluated against a public baseline on shared test sets. Recommended threshold: 0.50 — conservative default (EN precision 0.83 on the attack test, safe-prompt…

A fast, multilingual classifier that flags prompt injection and jailbreak attempts, both in user messages (direct) and in untrusted content an AI agent reads: emails, web pages, documents, RAG chunks, and tool/API outputs (indirect). their clean counterparts, so it looks for instructions aimed at the AI, not for scary words. - Low false-alarm rate on look-alike benign text: 89.7% on NotInject, 99.5% on OR-Bench-hard. half the size and the same decisions as fp32 on our checks), transformers.js. Labels: SAFE (0) and INJECTION (1). This is the same convention as protectai/deberta-v3-base-prompt-injection-v2, so the model is a drop-in replacement in code and tools built for that one. INJECTION…

Model · Zero-shot classification

proof_v3

Falcons.ai

A single-pass, typed, calibrated decision model. Give it a state (text, code or JSON), one or more typed questions (choice, noul yes/no, score ordinal) and a closed set of options. It returns a calibrated probability for every option, from one encoder pass per question, however many options there are. This repository is both the model card and the developer guide: how to load it, test it, and use it as a decision component in an agentic system. The model has no generative component. It can only rank the options you give it, so it cannot produce text outside that set. All numbers come from this checkpoint's falcondecreport.json: one run, small preset, seed 42, MODE="scratch". Calibration is…

These are chaeyidpt001/prot-asim-kr2-002 DreamBooth LoRA weights, trained on krea/Krea-2-Raw. The weights were trained using DreamBooth with the Krea 2 diffusers trainer. Krea 2 ships as two checkpoints: RAW (the non-distilled base you fine-tune on) and Turbo (an 8-step distilled checkpoint for fast, high-quality inference). Train your LoRA on RAW and run it on Turbo — LoRAs trained on RAW express strongly on Turbo. You should use TOK to trigger the image generation. Download the.safetensors LoRA in the Files & versions tab. For more details, including weighting, merging and fusing LoRAs, check the documentation on loading LoRAs in diffusers [TODO: provide examples of latent issues and…

Model

Protium-1.0

Solstice-AI

Protium-1.0 is Solstice-AI's high-efficiency, objective-gated cognitive reasoning architecture for open-weight foundation models. Named after Protium ($^1\text{H}$) — the fundamental isotope of hydrogen featuring zero neutrons and zero dead weight — Protium-1.0 replaces traditional, unconstrained chain-of-thought bloat with a lean, forward-only reasoning pipeline. Rather than imposing static token minimums, rigid pseudo-schemas, or unconstrained internal deliberation loops, Protium-1.0 operates on an Objective-Gated Protocol: every reasoning sequence anchors its verification criteria at step one, reasons strictly forward toward that objective, and terminates deliberation the moment verified…

This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.

本项目提供了 Psi0 算法在 Unitree G1 29 自由度教育版、BrainCo Revo1 触觉版右手与 DECXIN 摄像头组合上的训练和推理部署实现。 摄像头链接是设备参考，不作为性能验收依据；本仓库的图像规格以驱动和模型配置为准。BrainCo 的“触觉版”描述硬件型号，当前 Psi0 观测使用手部位置反馈，未将触觉阵列作为模型输入。 当前适配客户端从 WBC 的 ZMQ 遥测取得身体实测状态，WBC 自身通过 Unitree DDS 接收 G1 状态；客户端并非另开一条独立的 G1 DDS 状态链路。 相对原生 Psi0，本项目适配了 DECXIN 右目、BrainCo 状态与六维目标、跨机时间对齐、采集 initial token、推理启动门控及首段动作检查；保留 Psi0 模型服务和 SONIC token 控制的基本结构。 依赖清单记录安装状态，包含本地源码安装的包，不应将所有条目都视为可直接从 PyPI 安装。pc2 的上述两个补充包原先装在独立服务依赖目录中。 该安装脚本不启动训练，尚未在全新机器上完成安装验收。原 uv.lock 与实际安装版本记录应共同参考，详见 reproduction.json。 - 预训练 VLM、5k action-header 初始化权重，以及部署所用 40k 权重。 - 与训练一致的数据、冻结归一化统计及 initial token 文件。 - SONIC encoder、decoder、planner 权重和参考运动数据。 - WBC 构建依赖，包括 TensorRT、ONNX…

Model · Text generation

public-sft-qwen9b-v2

Alfaplus Laboratory

本仓库是一个 LoRA 适配器：在 Qwen/Qwen3.5-9B 上，用公开数据集 CohereLabs/ayadataset 的日文子集（languagecode == "jpn"）的 2400 条做了一轮指令微调（SFT）。 1. 几乎全部出自单一标注者：日文 original-annotations 共 6149 行，其中 6144 行出自同一个 userid（9881e959…），另一个标注者只有 5 行（整个日文 original-annotations 只有 2 个标注者）。这些内容里有大量「模板指令 + 一段やさしい日本語体的时事新闻（NHK NEWS WEB EASY 体，2022–2023 年）」，疑似搬自外部文章库而非标注者原创；用 WebSearch 对若干特征句做过核对，既未找到逐字一致的原文、也无法排除（相关旧文多已下线）。 2. 回答普遍很短：训练子集（2400 条）里 assistant 回答 54.2% 在 10 字以内、78.3% 在 50 字以内（多为问答题式的极简短答）。训练后模型很可能偏向极简短答。 3. 未做任何评测，且不得被当成「有日文能力」的模型：本轮只跑了训练，没有任何基准评测，也没有验证集；日文表现完全没有被测过。请不要把本适配器当成一个日文模型——它只是一次在公开日文数据上的小规模 SFT 实验。 不包含：合并后的完整权重、中间 checkpoint、优化器状态、训练语料或其副本。训练数据请从上游获取。 - 数据集：CohereLabs/ayadataset（Singh et al. 2024…

Qev-2B is the compact member of the Qev decision-model family. It learns option probabilities and representation responses from Qev-9B, and supports Choice, Noul and Score with the same request format. Install Qev from the source repository using Python 3.12 and a hardware-compatible PyTorch 2.8.0 build. The loader downloads this adaptation package and the compatible Qwen base separately. The teacher is not needed for inference. Use revision="v0.1.0" to select the first published release. To fine-tune on your own data, use configs/qev-2b-finetune.json with python -m qev.train --init-checkpoint AustinFu/Qev-2B. To train with a teacher, follow the distillation guide or 中文指南. Accuracy (%).…

Qev fine-tunes Qwen into a decision model. Give it context, a question, and answer options; receive a decision and a probability for each option. One model supports Choice, Noul (yes/no), and Score (ordered ratings). Qev-9B combines a general decision-training dataset with additional alignment and rule-compliance examples, repeated three times during the second half of training. It uses the full candidate-interaction architecture shown below. Use Python 3.12 and install a hardware-compatible PyTorch 2.8.0 build, then install Qev from its source repository: The Qev loader downloads this checkpoint and the pinned Qwen base separately. Use the Qev Python or JSONL interface to load the adapter…

static quants of https://huggingface.co/purav-raj321/Quantrexa-Prime-1.5B For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality…

Model · Text generation

quartz-micro-preview

Vertex AGI

The instruction-tuned release of Quartz Micro Preview Base — the same ~1B-parameter mixture-of-experts model, trained completely from scratch on a single consumer GPU, now given a full-parameter supervised fine-tune so it follows instructions instead of just completing text. This is a preview, not a finished product. It is small on every axis by design, and its evaluation below is reported honestly, including where it still falls short — that's this project's standing practice, not a disclaimer added after the fact. - Full-parameter SFT, not LoRA — the base model has zero instruction-following behavior, so a light adapter touch isn't the right tool; every parameter was updated.…

A 114,114,048-parameter language model trained from scratch on a single 8 GB laptop This is a base model. It continues text; it does not follow instructions or answer questions. At this size it writes fluent, on-topic prose that is often factually wrong. It is published as a baseline and as a record of what one laptop can train in a weekend, not as something to rely on. Validation loss on held-out FineWeb-Edu text, and on held-out code files deduplicated by exact hash against the training code. \ The code loss is flattered by the tokenizer. GPT-2's BPE splits indentation whitespace, against 2.0% for text. Those are easy to predict and pull the average down. The same effect makes greedy code…

Model · Text classification

Quorum-Reader-Qwen3-4B-Adapter

DataAgent

A LoRA adapter for Qwen/Qwen3-4B — the in-context reader of the Quorum 24-shot intent-classification ensemble. Given 24 retrieved examples in the prompt (no weight update), it scores every candidate label in a single forward pass. Combined (geometric mean) with a frozen BAAI/bge-large-en-v1.5 nearest-neighbour over the same 24 examples, the ensemble reaches 0.932 on Banking77 (full 3,080 test) — above the closed API Jev's 0.924 at Jev's own 24-shot protocol. In our study a further configuration (per-class descriptions + a calibration-selected combination) reached 0.938 and significantly beat Jev on a paired McNemar test LoRA (r = 64) on Qwen/Qwen3-4B, fine-tuned on a mixture of public…

Research checkpoints: Qwen/Qwen2.5-7B-Instruct fully fine-tuned on the ToolUse task with SFT or SDFT, on clean demonstrations or with 20% of demonstrations deliberately mismatched to other questions (mm20). Only the seed-42 final checkpoint of each of the 4 configurations is published; each lives in its own subfolder with a model card, sha256 list and eval results. mismatch models were trained on intentionally corrupted data and are for research only. PROTOCOL.md is the locked experiment protocol. manifests/ holds the train/val split (record ids) and the corruption mapping (record indices + hashes; no dataset text).

A high-throughput inference engine for structured information extraction, decision routing, and categorical classification on Apple Silicon using MLX. Parallel Constrained Decoding evaluates multi-field JSON schemas simultaneously rather than generating tokens sequentially. On an Apple Silicon M4 Max, it delivers 5.6x to 7.0x latency reductions compared to standard autoregressive decoding with 100% schema validity and calibrated field-level confidence scores. Evaluated with mlx-community/Qwen2.5-1.5B-Instruct-4bit on macOS Sequoia: Standard LLM structured generation (such as JSON mode or grammar-guided sampling) relies on token-by-token autoregressive decoding: Each token requires a…

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Model · Text generation

Qwen-3.8-9B-int4-ov

Asahi

This is an OpenVINO INT4 quantized version of empero-ai/Qwen3.8-9B-Distill, generated using optimum-intel. For detailed specifications, evaluation scores, and architecture info, please refer to the original repository: empero-ai/Qwen3.8-9B-Distill.

Model · Text generation

Qwen-3.8-9B-int8-ov

Asahi

This is an OpenVINO INT8 quantized version of empero-ai/Qwen3.8-9B-Distill, generated using optimum-intel. For detailed specifications, evaluation scores, and architecture info, please refer to the original repository: empero-ai/Qwen3.8-9B-Distill.

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Model · Image to image

Qwen-Image-2.1-Multiple-Angles-LoRA

AK

Reshoot any object from a controlled camera position: feed a photo, ask for a different azimuth/elevation, get the same subject from that angle. First camera-control LoRA for Qwen-Image-2.1 — the newest Qwen image base (unified generate + edit, native RGBA), where this format did not exist yet. - 36 poses: 12 azimuths (30° steps) × 4 elevations (0° eye-level / 30° elevated / 60° high-angle / 90° top-down), plus a close-up variant of each → 72 addressable framings In-distribution examples are above — but does it work on everyday photos it never saw? Dog, cat and apple from Wikimedia Commons, step-1,000 checkpoint, same prompt grammar: Golden retriever · front-right quarter view, eye-level…

Model · Image to image

Qwen-Image-2.1-Next-Scene-LoRA

AK

Feed a frame, direct the next shot. This LoRA teaches Qwen-Image-2.1 cinematic scene continuation: it takes an input image and a director-style instruction and generates the next scene — camera moves, reveals, lighting shifts — while keeping the world, characters and style consistent. Chain outputs into storyboards. This is the Qwen-Image-2.1 port of the next-scene format pioneered by (313K downloads on the 2509 edit model) — retrained natively on the newest Qwen image base, where this capability did not exist yet. Companion piece to (camera control) — together they cover where the camera is and where the story goes next. ComfyUI: load in any Qwen-Image-2.1 workflow with a LoRA node…