vlm-segmentation
Skill213 starsChanged 2 days agoMITView on GitHub
Choose and evaluate VLM or segmentation pipelines, including text-conditioned detection, masks, part labels, model-license constraints, and measured GPU deployment choices. Use when a task has a VLM or segmentation component; route pure diffusion prompting, training, or serving to its specialized skill.
Files
VLM + Segmentation + Diffusion Engineering
Скилл охватывает три тесно связанных домена. Выбери нужный раздел и загрузи соответствующий reference-файл.
Навигация по доменам
| Задача | Reference файл |
| Выбор модели сегментации, pipeline "текст → маски", VLM-стек, part-labeling | references/vlm-segmentation.md |
| Диффузионные архитектуры, schedulers, обучение, LoRA, text encoder fusion | references/diffusion-engineering.md |
| Два инстанса SAM3 на H100, MIG/MPS, memory, профилирование | references/gpu-deployment.md |
Правило выбора: если вопрос смешивает темы (например, "как деплоить диффузионную модель на H100") — прочитай оба релевантных файла.
Быстрые ответы без чтения reference-файлов
Candidate pipeline "фраза → маски"
1. SAM3 PCS (текстовый концепт) → instance masks + boxes + scores
ИЛИ
Grounding DINO / OWLv2 / YOLO-World → boxes → SAM2.1 → masks
2. Part-labeling: отдельный классификатор по ROI + фиксированный словарьCandidate diffusion pipeline
1. Backbone: UNet (просто) или DiT/Flow (масштабирование)
2. Latent diffusion (VAE → латенты → денойзер → VAE decode)
3. Text encoder: CLIP (SD), два CLIP (SDXL), Qwen3 (Flux.2 klein 9B)
4. Fine-tune: начинать с LoRA, full fine-tune только если нужно
5. Memory: AMP (BF16) → checkpointing → ZeRO/FSDP при масштабеTwo SAM3 instances on H100 (only after host inspection)
MIG can provide hardware partitioning where the inspected GPU, driver, current
MIG layout and workload support it. It changes host GPU configuration: preserve
the current layout and obtain explicit operational approval before any change.
MPS (fallback) → кооперативный шеринг, без строгой изоляцииКлючевые характеристики моделей (быстрая справка)
| Модель | Параметры | Лицензия | Главная сильная сторона |
| SAM3 | 848M | SAM License (gated) | Open-vocab сегментация по тексту, все инстансы |
| SAM2.1-large | model-card specific | Apache-2.0 | Видео-трекинг, интерактивная сегментация; reproduce any FPS on the target stack |
| SAM2.1-tiny | model-card specific | Apache-2.0 | Lightweight variant; reproduce any FPS on the target stack |
| Florence-2-large | 770M | MIT | Унифицированные задачи через task prompt |
| EdgeTAM | ~SAM2-tiny | Apache-2.0 | 16 FPS на iPhone 15 Pro Max, CoreML |
| Grounding DINO | — | Apache-2.0 | Text-conditioned detection, boxes |
| YOLO-World | — | GPL-3.0 | Real-time open-vocab OD, 52 FPS V100 |
Критические предупреждения
- SAM3: gated access на HF, кастомная SAM License — проверь перед продакшном
- YOLO-World: upstream states GPL-3.0 and supports commercial usage. GPL obligations apply by default; obtain a separate commercial licence only when the intended distribution or policy requires terms outside GPL, with legal review for the specific product.
- Замена text encoder: не plug-and-play, нужен projection + переобучение cross-attention
- MIG vs MPS: только MIG даёт аппаратную изоляцию VRAM/SM; MPS — кооперативный шеринг
- For non-English prompts, compare the target model’s supported languages on a representative evaluation set; do not silently translate or claim a universal English advantage.
Other versions1 project uses a version of this.
Word for word · 1
- AnastasiyaW/claude-code-config★ 209100% same
Edited · 0
None found.
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository60 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- agent-harness-designSkill
- agents-best-practicesSkill
- diffusion-engineeringSkill
- flux2-klein-promptingSkill
- flux2-lora-trainingSkill
- forensic-prompt-compilerSkill
- ml-research-labSkill
- notebooklm-grounded-researchSkill
- feature-newSkill
- harness-designSkill
- layer-newSkill
- plan-swarm-reviewSkill
- pixel-art-storyboardSkill
- pixel-art-studioSkill
- architecture-firstSkill
- architecture-qualitySkill
- bug-reproducerSkill
- code-complexitySkill
- control-cliSkill
- control-uiSkill
- deep-reviewSkill
- deslopSkill
- distill-feedbackSkill
- harness-feedbackSkill
- proof-verifySkill
- refactoring-safelySkill
- repo-mapSkill
- root-cause-deliverySkill
- system-and-data-designSkill
- testing-strategySkill
- thermo-nuclear-code-quality-reviewSkill
- verify-thisSkill
- workflow-orchestrationSkill
- frontend-designSkill
- motion-framerSkill
- ui-designSkill
- ui-ux-pro-maxSkill
- ios-developmentSkill
- lean-codeSkill
- cross-harness-continuationSkill
- deepseek-provider-contractSkill
- desktop-sessions-discoverySkill
- epistemic-challengeSkill
- gemini-delegateSkill
- harness-auditSkill
- observability-monitoringSkill
- remote-compute-opsSkill
- plan-to-ticketsSkill
- product-meaning-extractorSkill
- remotion-production-guideSkill
- script-evaluatorSkill
- video-narrative-arcSkill
- video-post-productionSkill
- article-structure-reviewSkill
- cavemanSkill
- humanize-englishSkill
- humanize-russianSkill
- owner-voiceSkill
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

