XB-BOX - 采样器（原版优化）

This is not a new sampler. Read the source and you'll see it's a straight wrapper around ComfyUI's own KSampler - same inputs, same outputs, same sampling code, called "一字不差" (word for word, per the author). What XB_KSampler adds is exactly one thing: a cleanup dropdown that runs before sampling. On a VRAM-starved card, that's genuinely worth having.

If you're running a big video model or stacking a VAE decode right before sampling, PyTorch's cached allocator can leave you with a fragmented, half-empty VRAM pool - plenty of "used" memory that's actually just cached blobs waiting to be freed. The stock KSampler doesn't care; it just tries to allocate and can OOM for no good reason. This node lets you force a clean-up pass first.

The cleanup dropdown

Four options, escalating from nothing to nuclear:

• 不做任何清理 (no cleanup) - the default, and it's just the stock sampler.
• 单次缓存清理 (single cache cleanup) - soft_empty_cache() + empty_cache(), clearing the PyTorch cache without unloading anything. Cheap, fast, the one you'll reach for most.
• 卸载显存模型 (unload VRAM models) - the above plus unload_all_models(), which pushes models back to system RAM before sampling. Costs a reload, buys real headroom.
• 卸载全量模型 (full cleanup) - everything above plus cleanup_models(), gc.collect(), and IPC collect. Slowest, freest VRAM.

Everything else

Every other input is identical to the vanilla KSampler: model, seed, steps, cfg, sampler_name (the full 44-option list), scheduler, positive, negative, latent_image, denoise. Same single LATENT output, same behavior. The sampler theory you already know applies unchanged - on flow-matching models (Flux, Wan, LTX) stick to euler-family samplers with conservative schedulers; the Karras habit from SDXL doesn't carry over. The wrapper adds nothing to that equation.

The whole point of this node is the workflow ergonomics: instead of threading a VRAM Cleaner node into the graph before your sampler, you flip a dropdown on the sampler itself. It also guards against the classic low-VRAM failure where a big VAE decode right before sampling leaves the cache full and the sampler OOMs on a perfectly reasonable step count.

Installing it

Standard XB_ToolBox install:

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git

or via ComfyUI Manager (search XB_ToolBox), then restart. No extra dependencies for this node - it only touches ComfyUI core.

When to actually use it

If you never hit VRAM errors, leave the dropdown on no-cleanup and you've effectively got the stock sampler with a cosmetic rename. If you do hit intermittent OOMs at sampling time, especially on 8GB cards with video models, set it to single cache cleanup and see if the problem disappears - if it doesn't, escalate to unload-models. The catch is that aggressive cleanup costs time: unloading and reloading models on every single generation adds measurable latency, so the full option is really for "one-shot bake" workflows, not daily interactive tweaking. And note it runs its cleanup before sampling only - it doesn't help the VAE decode on the way out.