在進行 AI 影片生成時，硬體資源的使用狀況往往直接影響生成速度與穩定性，尤其是對 VRAM 與 RAM 需求相對敏感的模型而言，更需要即時掌握系統負載。WanGP 作為一套以低 VRAM […]

隨著開源影片生成技術逐漸成熟，即使在 GPU 資源有限的情況下，也能實現具備高完成度的角色動畫作品。WanGP 正是為此而生的整合型平台，透過友善的 Web 介面與高度模組化的設計， […]

近年來，隨著生成式 AI 與擴散模型的快速演進，影音內容的自動化生成已不再侷限於靜態圖片或簡單動畫，而是逐步邁向「可控、可對齊、可延展」的動態影片層級。特別是在語音驅動角色動畫（Audio-dr […]

在生成式 AI 影片快速演進的這幾年，「本地端」重新成為創作者關注的核心。雲端平台雖然方便，卻伴隨著費用、速度與隱私等問題；對於希望掌控整個創作流程的使用者而言，能否在自己的電腦上，完整跑起模型、 […]

Hmm… it looks like there are a few places where this could be getting stuck, so I think a quick bit of isolation would probably get to the fix faster: * * * I don’t think the log is enough to conclude **“MiniMax H3 simply does not fit on this GPU”** yet. There is actually a fairly useful clue in the traceback: the failure is happening in the **FL2VA Start Image conditioning path** , around the point where WanGP prepares the first image for the H3 Video VAE: generate -> _add_image_condition(image_start, ...) -> _encode_video(video) -> video.unsqueeze(0).to(cuda, ...) -> vae.encode_condition(...) That is a substantially narrower starting point than a generic OOM during denoising. Also, the fact that Wan 2.2 works is useful as a control — it suggests the CUDA installation is broadly functional — but it does **not** prove that H3’s particular component/offload path has enough memory headroom. H3 has a rather different memory layout, and other H3 backends also have to treat its text encoder, DiT and Video VAE as separate memory-management stages. So before changing the VAE code or a bunch of memory settings at once, I would try two very small controls. ## 1. Same H3 settings, but remove only the Start Image Try the same: * H3 FL2VA checkpoint * prompt * memory profile * duration / frame count * **explicit output width and height** but remove only the Start Image. The explicit canvas is important: H3 can derive its default canvas/aspect ratio from the first keyframe, so removing the image without fixing width/height may accidentally change two variables at once. The Diffusers H3 documentation describes the same distinction between text-only `t2va` and first/last-keyframe `fl2va`. The result would separate two fairly different cases: same H3 + same explicit canvas | remove Start Image | +------+------+ | | works OOM | | v v Start Image / H3-wide Video VAE memory/runtime conditioning issue still path becomes in play interesting If **H3 works without the Start Image** , I would not start changing the transformer, attention backend, or Qwen quantization yet. The failure has become much more specifically connected to image conditioning. If **H3 still OOMs without the Start Image** , then I would broaden the diagnosis back out to GPU/VRAM, system RAM, MMGP profile, canvas size, duration, etc. ## 2. If removing the Start Image fixes it, try one completely fresh process Fully exit WanGP/Pinokio’s running WanGP process, start it again, and make the problematic **H3 + Start Image** job the first generation. If that first run succeeds, repeat the exact same generation once. fresh WanGP process | H3 + Start Image | +---+---+ | | works OOM | | repeat stale state from a once previous generation | becomes much less | interesting | +-- second run OOM | v retained state / component residency / lifecycle becomes much more interesting That is useful because the current H3 pipeline processes the Start Image **before** its normal prompt/Qwen encoding stage. In other words, the simple explanation “Qwen and the H3 transformer were already loaded by this same generation and then crowded out the VAE” does not match the normal ordering very well. If only a later generation fails, however, then retained state becomes much more plausible. If both fresh and repeated runs fail identically, I would look elsewhere. ## The few environment details that would make the next branch much clearer If you post another result, I think these are enough; a giant environment dump probably is not necessary yet: * exact GPU model and VRAM * physical system RAM * exact WanGP version / update level * MMGP profile (1–5 / 4+ if applicable) * output width × height * duration or frame count * whether the same H3 job works with no Start Image * optionally, free VRAM immediately before pressing Generate On Windows, a simple `nvidia-smi` immediately before the run is already useful. WanGP itself also notes that a GPU-accelerated web browser can consume a surprisingly large amount of VRAM and even ships CPU-only Chrome launch scripts for this reason, so closing unrelated GPU applications/browser windows is a reasonable cheap sanity check — not a diagnosis, just a way to remove one variable. See the WanGP README. So, for now, my rough interpretation would be: Observation | What I would investigate next ---|--- H3 works with fixed canvas but no Start Image | Start Image / Video VAE conditioning boundary H3 OOMs with or without the image | H3-wide VRAM/profile/hardware conditions Fresh first run works, later identical run fails | retained GPU state / lifecycle / residency Fresh first run also fails | deterministic memory/path issue rather than stale previous-run state GPU is already nearly full before Generate | external/unexpected GPU residency Synchronized traceback moves deeper into VAE encode | VAE encoder working-set / load peak Why the traceback is more interesting than a generic OOM (click for more details) Why I would not treat the `video_encoder` MMGP budget as the first suspect (click for more details) How I would read the MMGP partial-pinning messages (click for more details) Why Wan 2.2 working does not rule out an H3-specific memory path (click for more details) There are some similar H3 reports, but I would not assume they have the same cause (click for more details) If the two simple controls still leave the failure ambiguous (click for more details) Things I would not change yet (click for more details) So, if I were trying to minimize the amount of experimentation, the most useful next result would simply be one of these: A. Same explicit canvas, H3 works when Start Image is removed. B. Same explicit canvas, H3 still OOMs without Start Image. C. Start Image works on the first run after a full WanGP restart, but the same job fails on a later run. D. Start Image fails even on the first clean run. Then add: GPU: VRAM: RAM: WanGP version: MMGP profile: Width x height: Duration / frames: Those outcomes point in fairly different directions. **A** would make me inspect the Start Image / H3 Video Encoder boundary next. **B** would make me look much harder at overall H3 memory budget/profile/hardware conditions. **C** would make process state/model residency considerably more suspicious. **D** , especially if the GPU appears to have plenty of free memory immediately before Generate, would make one `CUDA_LAUNCH_BLOCKING=1` diagnostic run quite worthwhile. At this point I think those controls will tell us substantially more than changing several memory settings or editing the VAE code at random. One small privacy note if you decide to attach debugging material: WanGP’s queue ZIP format can contain embedded image/video/audio attachments, not just settings. I would inspect any `error_queue.zip` before posting it publicly; if settings are all that are needed, a redacted settings/JSON dump is safer.

Hmm… it looks like there are a few places where this could be getting stuck, so I think a quick bit of isolation would probably get to the fix faster: * * * I don’t think the log is enough to conclude **“MiniMax H3 simply does not fit on this GPU”** yet. There is actually a fairly useful clue in the traceback: the failure is happening in the **FL2VA Start Image conditioning path** , around the point where WanGP prepares the first image for the H3 Video VAE: generate -> _add_image_condition(image_start, ...) -> _encode_video(video) -> video.unsqueeze(0).to(cuda, ...) -> vae.encode_condition(...) That is a substantially narrower starting point than a generic OOM during denoising. Also, the fact that Wan 2.2 works is useful as a control — it suggests the CUDA installation is broadly functional — but it does **not** prove that H3’s particular component/offload path has enough memory headroom. H3 has a rather different memory layout, and other H3 backends also have to treat its text encoder, DiT and Video VAE as separate memory-management stages. So before changing the VAE code or a bunch of memory settings at once, I would try two very small controls. ## 1. Same H3 settings, but remove only the Start Image Try the same: * H3 FL2VA checkpoint * prompt * memory profile * duration / frame count * **explicit output width and height** but remove only the Start Image. The explicit canvas is important: H3 can derive its default canvas/aspect ratio from the first keyframe, so removing the image without fixing width/height may accidentally change two variables at once. The Diffusers H3 documentation describes the same distinction between text-only `t2va` and first/last-keyframe `fl2va`. The result would separate two fairly different cases: same H3 + same explicit canvas | remove Start Image | +------+------+ | | works OOM | | v v Start Image / H3-wide Video VAE memory/runtime conditioning issue still path becomes in play interesting If **H3 works without the Start Image** , I would not start changing the transformer, attention backend, or Qwen quantization yet. The failure has become much more specifically connected to image conditioning. If **H3 still OOMs without the Start Image** , then I would broaden the diagnosis back out to GPU/VRAM, system RAM, MMGP profile, canvas size, duration, etc. ## 2. If removing the Start Image fixes it, try one completely fresh process Fully exit WanGP/Pinokio’s running WanGP process, start it again, and make the problematic **H3 + Start Image** job the first generation. If that first run succeeds, repeat the exact same generation once. fresh WanGP process | H3 + Start Image | +---+---+ | | works OOM | | repeat stale state from a once previous generation | becomes much less | interesting | +-- second run OOM | v retained state / component residency / lifecycle becomes much more interesting That is useful because the current H3 pipeline processes the Start Image **before** its normal prompt/Qwen encoding stage. In other words, the simple explanation “Qwen and the H3 transformer were already loaded by this same generation and then crowded out the VAE” does not match the normal ordering very well. If only a later generation fails, however, then retained state becomes much more plausible. If both fresh and repeated runs fail identically, I would look elsewhere. ## The few environment details that would make the next branch much clearer If you post another result, I think these are enough; a giant environment dump probably is not necessary yet: * exact GPU model and VRAM * physical system RAM * exact WanGP version / update level * MMGP profile (1–5 / 4+ if applicable) * output width × height * duration or frame count * whether the same H3 job works with no Start Image * optionally, free VRAM immediately before pressing Generate On Windows, a simple `nvidia-smi` immediately before the run is already useful. WanGP itself also notes that a GPU-accelerated web browser can consume a surprisingly large amount of VRAM and even ships CPU-only Chrome launch scripts for this reason, so closing unrelated GPU applications/browser windows is a reasonable cheap sanity check — not a diagnosis, just a way to remove one variable. See the WanGP README. So, for now, my rough interpretation would be: Observation | What I would investigate next ---|--- H3 works with fixed canvas but no Start Image | Start Image / Video VAE conditioning boundary H3 OOMs with or without the image | H3-wide VRAM/profile/hardware conditions Fresh first run works, later identical run fails | retained GPU state / lifecycle / residency Fresh first run also fails | deterministic memory/path issue rather than stale previous-run state GPU is already nearly full before Generate | external/unexpected GPU residency Synchronized traceback moves deeper into VAE encode | VAE encoder working-set / load peak Why the traceback is more interesting than a generic OOM (click for more details) Why I would not treat the `video_encoder` MMGP budget as the first suspect (click for more details) How I would read the MMGP partial-pinning messages (click for more details) Why Wan 2.2 working does not rule out an H3-specific memory path (click for more details) There are some similar H3 reports, but I would not assume they have the same cause (click for more details) If the two simple controls still leave the failure ambiguous (click for more details) Things I would not change yet (click for more details) So, if I were trying to minimize the amount of experimentation, the most useful next result would simply be one of these: A. Same explicit canvas, H3 works when Start Image is removed. B. Same explicit canvas, H3 still OOMs without Start Image. C. Start Image works on the first run after a full WanGP restart, but the same job fails on a later run. D. Start Image fails even on the first clean run. Then add: GPU: VRAM: RAM: WanGP version: MMGP profile: Width x height: Duration / frames: Those outcomes point in fairly different directions. **A** would make me inspect the Start Image / H3 Video Encoder boundary next. **B** would make me look much harder at overall H3 memory budget/profile/hardware conditions. **C** would make process state/model residency considerably more suspicious. **D** , especially if the GPU appears to have plenty of free memory immediately before Generate, would make one `CUDA_LAUNCH_BLOCKING=1` diagnostic run quite worthwhile. At this point I think those controls will tell us substantially more than changing several memory settings or editing the VAE code at random. One small privacy note if you decide to attach debugging material: WanGP’s queue ZIP format can contain embedded image/video/audio attachments, not just settings. I would inspect any `error_queue.zip` before posting it publicly; if settings are all that are needed, a redacted settings/JSON dump is safer.

Hmm… it looks like there are a few places where this could be getting stuck, so I think a quick bit of isolation would probably get to the fix faster: * * * I don’t think the log is enough to conclude **“MiniMax H3 simply does not fit on this GPU”** yet. There is actually a fairly useful clue in the traceback: the failure is happening in the **FL2VA Start Image conditioning path** , around the point where WanGP prepares the first image for the H3 Video VAE: generate -> _add_image_condition(image_start, ...) -> _encode_video(video) -> video.unsqueeze(0).to(cuda, ...) -> vae.encode_condition(...) That is a substantially narrower starting point than a generic OOM during denoising. Also, the fact that Wan 2.2 works is useful as a control — it suggests the CUDA installation is broadly functional — but it does **not** prove that H3’s particular component/offload path has enough memory headroom. H3 has a rather different memory layout, and other H3 backends also have to treat its text encoder, DiT and Video VAE as separate memory-management stages. So before changing the VAE code or a bunch of memory settings at once, I would try two very small controls. ## 1. Same H3 settings, but remove only the Start Image Try the same: * H3 FL2VA checkpoint * prompt * memory profile * duration / frame count * **explicit output width and height** but remove only the Start Image. The explicit canvas is important: H3 can derive its default canvas/aspect ratio from the first keyframe, so removing the image without fixing width/height may accidentally change two variables at once. The Diffusers H3 documentation describes the same distinction between text-only `t2va` and first/last-keyframe `fl2va`. The result would separate two fairly different cases: same H3 + same explicit canvas | remove Start Image | +------+------+ | | works OOM | | v v Start Image / H3-wide Video VAE memory/runtime conditioning issue still path becomes in play interesting If **H3 works without the Start Image** , I would not start changing the transformer, attention backend, or Qwen quantization yet. The failure has become much more specifically connected to image conditioning. If **H3 still OOMs without the Start Image** , then I would broaden the diagnosis back out to GPU/VRAM, system RAM, MMGP profile, canvas size, duration, etc. ## 2. If removing the Start Image fixes it, try one completely fresh process Fully exit WanGP/Pinokio’s running WanGP process, start it again, and make the problematic **H3 + Start Image** job the first generation. If that first run succeeds, repeat the exact same generation once. fresh WanGP process | H3 + Start Image | +---+---+ | | works OOM | | repeat stale state from a once previous generation | becomes much less | interesting | +-- second run OOM | v retained state / component residency / lifecycle becomes much more interesting That is useful because the current H3 pipeline processes the Start Image **before** its normal prompt/Qwen encoding stage. In other words, the simple explanation “Qwen and the H3 transformer were already loaded by this same generation and then crowded out the VAE” does not match the normal ordering very well. If only a later generation fails, however, then retained state becomes much more plausible. If both fresh and repeated runs fail identically, I would look elsewhere. ## The few environment details that would make the next branch much clearer If you post another result, I think these are enough; a giant environment dump probably is not necessary yet: * exact GPU model and VRAM * physical system RAM * exact WanGP version / update level * MMGP profile (1–5 / 4+ if applicable) * output width × height * duration or frame count * whether the same H3 job works with no Start Image * optionally, free VRAM immediately before pressing Generate On Windows, a simple `nvidia-smi` immediately before the run is already useful. WanGP itself also notes that a GPU-accelerated web browser can consume a surprisingly large amount of VRAM and even ships CPU-only Chrome launch scripts for this reason, so closing unrelated GPU applications/browser windows is a reasonable cheap sanity check — not a diagnosis, just a way to remove one variable. See the WanGP README. So, for now, my rough interpretation would be: Observation | What I would investigate next ---|--- H3 works with fixed canvas but no Start Image | Start Image / Video VAE conditioning boundary H3 OOMs with or without the image | H3-wide VRAM/profile/hardware conditions Fresh first run works, later identical run fails | retained GPU state / lifecycle / residency Fresh first run also fails | deterministic memory/path issue rather than stale previous-run state GPU is already nearly full before Generate | external/unexpected GPU residency Synchronized traceback moves deeper into VAE encode | VAE encoder working-set / load peak Why the traceback is more interesting than a generic OOM (click for more details) Why I would not treat the `video_encoder` MMGP budget as the first suspect (click for more details) How I would read the MMGP partial-pinning messages (click for more details) Why Wan 2.2 working does not rule out an H3-specific memory path (click for more details) There are some similar H3 reports, but I would not assume they have the same cause (click for more details) If the two simple controls still leave the failure ambiguous (click for more details) Things I would not change yet (click for more details) So, if I were trying to minimize the amount of experimentation, the most useful next result would simply be one of these: A. Same explicit canvas, H3 works when Start Image is removed. B. Same explicit canvas, H3 still OOMs without Start Image. C. Start Image works on the first run after a full WanGP restart, but the same job fails on a later run. D. Start Image fails even on the first clean run. Then add: GPU: VRAM: RAM: WanGP version: MMGP profile: Width x height: Duration / frames: Those outcomes point in fairly different directions. **A** would make me inspect the Start Image / H3 Video Encoder boundary next. **B** would make me look much harder at overall H3 memory budget/profile/hardware conditions. **C** would make process state/model residency considerably more suspicious. **D** , especially if the GPU appears to have plenty of free memory immediately before Generate, would make one `CUDA_LAUNCH_BLOCKING=1` diagnostic run quite worthwhile. At this point I think those controls will tell us substantially more than changing several memory settings or editing the VAE code at random. One small privacy note if you decide to attach debugging material: WanGP’s queue ZIP format can contain embedded image/video/audio attachments, not just settings. I would inspect any `error_queue.zip` before posting it publicly; if settings are all that are needed, a redacted settings/JSON dump is safer.

Hmm… it looks like there are a few places where this could be getting stuck, so I think a quick bit of isolation would probably get to the fix faster: * * * I don’t think the log is enough to conclude **“MiniMax H3 simply does not fit on this GPU”** yet. There is actually a fairly useful clue in the traceback: the failure is happening in the **FL2VA Start Image conditioning path** , around the point where WanGP prepares the first image for the H3 Video VAE: generate -> _add_image_condition(image_start, ...) -> _encode_video(video) -> video.unsqueeze(0).to(cuda, ...) -> vae.encode_condition(...) That is a substantially narrower starting point than a generic OOM during denoising. Also, the fact that Wan 2.2 works is useful as a control — it suggests the CUDA installation is broadly functional — but it does **not** prove that H3’s particular component/offload path has enough memory headroom. H3 has a rather different memory layout, and other H3 backends also have to treat its text encoder, DiT and Video VAE as separate memory-management stages. So before changing the VAE code or a bunch of memory settings at once, I would try two very small controls. ## 1. Same H3 settings, but remove only the Start Image Try the same: * H3 FL2VA checkpoint * prompt * memory profile * duration / frame count * **explicit output width and height** but remove only the Start Image. The explicit canvas is important: H3 can derive its default canvas/aspect ratio from the first keyframe, so removing the image without fixing width/height may accidentally change two variables at once. The Diffusers H3 documentation describes the same distinction between text-only `t2va` and first/last-keyframe `fl2va`. The result would separate two fairly different cases: same H3 + same explicit canvas | remove Start Image | +------+------+ | | works OOM | | v v Start Image / H3-wide Video VAE memory/runtime conditioning issue still path becomes in play interesting If **H3 works without the Start Image** , I would not start changing the transformer, attention backend, or Qwen quantization yet. The failure has become much more specifically connected to image conditioning. If **H3 still OOMs without the Start Image** , then I would broaden the diagnosis back out to GPU/VRAM, system RAM, MMGP profile, canvas size, duration, etc. ## 2. If removing the Start Image fixes it, try one completely fresh process Fully exit WanGP/Pinokio’s running WanGP process, start it again, and make the problematic **H3 + Start Image** job the first generation. If that first run succeeds, repeat the exact same generation once. fresh WanGP process | H3 + Start Image | +---+---+ | | works OOM | | repeat stale state from a once previous generation | becomes much less | interesting | +-- second run OOM | v retained state / component residency / lifecycle becomes much more interesting That is useful because the current H3 pipeline processes the Start Image **before** its normal prompt/Qwen encoding stage. In other words, the simple explanation “Qwen and the H3 transformer were already loaded by this same generation and then crowded out the VAE” does not match the normal ordering very well. If only a later generation fails, however, then retained state becomes much more plausible. If both fresh and repeated runs fail identically, I would look elsewhere. ## The few environment details that would make the next branch much clearer If you post another result, I think these are enough; a giant environment dump probably is not necessary yet: * exact GPU model and VRAM * physical system RAM * exact WanGP version / update level * MMGP profile (1–5 / 4+ if applicable) * output width × height * duration or frame count * whether the same H3 job works with no Start Image * optionally, free VRAM immediately before pressing Generate On Windows, a simple `nvidia-smi` immediately before the run is already useful. WanGP itself also notes that a GPU-accelerated web browser can consume a surprisingly large amount of VRAM and even ships CPU-only Chrome launch scripts for this reason, so closing unrelated GPU applications/browser windows is a reasonable cheap sanity check — not a diagnosis, just a way to remove one variable. See the WanGP README. So, for now, my rough interpretation would be: Observation | What I would investigate next ---|--- H3 works with fixed canvas but no Start Image | Start Image / Video VAE conditioning boundary H3 OOMs with or without the image | H3-wide VRAM/profile/hardware conditions Fresh first run works, later identical run fails | retained GPU state / lifecycle / residency Fresh first run also fails | deterministic memory/path issue rather than stale previous-run state GPU is already nearly full before Generate | external/unexpected GPU residency Synchronized traceback moves deeper into VAE encode | VAE encoder working-set / load peak Why the traceback is more interesting than a generic OOM (click for more details) Why I would not treat the `video_encoder` MMGP budget as the first suspect (click for more details) How I would read the MMGP partial-pinning messages (click for more details) Why Wan 2.2 working does not rule out an H3-specific memory path (click for more details) There are some similar H3 reports, but I would not assume they have the same cause (click for more details) If the two simple controls still leave the failure ambiguous (click for more details) Things I would not change yet (click for more details) So, if I were trying to minimize the amount of experimentation, the most useful next result would simply be one of these: A. Same explicit canvas, H3 works when Start Image is removed. B. Same explicit canvas, H3 still OOMs without Start Image. C. Start Image works on the first run after a full WanGP restart, but the same job fails on a later run. D. Start Image fails even on the first clean run. Then add: GPU: VRAM: RAM: WanGP version: MMGP profile: Width x height: Duration / frames: Those outcomes point in fairly different directions. **A** would make me inspect the Start Image / H3 Video Encoder boundary next. **B** would make me look much harder at overall H3 memory budget/profile/hardware conditions. **C** would make process state/model residency considerably more suspicious. **D** , especially if the GPU appears to have plenty of free memory immediately before Generate, would make one `CUDA_LAUNCH_BLOCKING=1` diagnostic run quite worthwhile. At this point I think those controls will tell us substantially more than changing several memory settings or editing the VAE code at random. One small privacy note if you decide to attach debugging material: WanGP’s queue ZIP format can contain embedded image/video/audio attachments, not just settings. I would inspect any `error_queue.zip` before posting it publicly; if settings are all that are needed, a redacted settings/JSON dump is safer.

YouTube video by Fuzz Puppy

Google Colabで動作する動画生成AI「Wan 2.2 Animate (WanGP)」の使い方を徹底解説。Video-to-Video技術で実写動画の人物をアニメや別キャラクターに置換する方法、マスク機能の活用、LoRA設定まで、Python環境構築なしで試せる手順を紹介します。<p>Copyright &copy; 2026 <a href="https://okihiro-school.com">OKIHIRO AI Creative</a> All Rights Reserved.</p>

話題の動画生成AI「LTX-2」をローカル環境で動かす方法を解説。低VRAMでも動作する最適化ツール「WanGP」のインストールから使い方まで、Pythonエンジニア向けに詳しく紹介します。<p>Copyright &copy; 2026 <a href="https://okihiro-school.com">OKIHIRO AI Creative</a> All Rights Reserved.</p>

Wan 2.2 I2Vの新機能「First Frame & Last Frame」を使いこなし、動画のストーリーを完全に制御する方法を解説。ComfyUIとWanGP、そしてQwen Image Edit用「Next Scene LoRA」を組み合わせた最新ワークフローです。<p>Copyright &copy; 2025 <a href="https://okihiro-school.com">OKIHIRO AI Creative</a> All Rights Reserved.</p>

YouTube video by Social&Apps