IAMCCS Detail Atelier

Here's the honest pitch: LTX-2 video looks good, but "looks good" isn't "looks finished." A first pass at 768px with 24 frames has soft faces and mushy textures, and upscaling it after the fact just upscales the mush. The Detail Atelier is IAMCCS's answer - a dedicated detail-enhancement sampler that re-runs the sampling pass on your video latents with tiling, guiding, and VRAM presets tuned for consumer cards. It's the difference between "I generated a video" and "I'd show this to a client."

The catch: it's a custom sampler, which means you wire it up like SamplerCustomAdvanced, not like a checkbox node. You bring your own model, VAE, noise, sampler, sigmas, and guider. The Atelier handles everything hard - the tiling, the VRAM juggling, the re-encoding.

How it works

At its core it runs the same sampling loop ComfyUI's advanced sampler does, but with three tricks layered on:

• Temporal + spatial tiling. Long clips are cut into overlapping temporal tiles so each pass fits in VRAM; optional spatial tiling (horizontal_tiles/vertical_tiles via the Advanced config) splits wide frames too. Overlaps keep the seams invisible.
• Guiding. guiding_start_step/guiding_end_step control when in the schedule the detail guidance kicks in, and the optional optional_cond_images / optional_guiding_latents inputs anchor specific frames or shots so the detail pass doesn't wander off-identity.
• VRAM presets. Pick vram_preset (8GB/12GB/16GB/24GB) and quality_mode (preview/balanced/quality) and the node picks tile sizes, overlaps, and target resolution for you. The presets are real and hand-tuned - for example 12GB/balanced runs a 1280 long-edge pass with 56-frame temporal tiles and 24-frame overlap.

You also get adain_factor (style-transfer-ish normalization between tiles - keep it at 0 until you understand what it does), oom_retry (re-attempts on OOM, worth leaving on), and hardware_mode auto/manual. Outputs are denoised_output (LATENT), ready to VAE-decode, plus report_json and report.

The inputs that matter

• model, vae, noise, sampler, sigmas, guider, latents - the standard custom-sampler wiring. The tooltips spell each out; latents is the video you want to detail.
• vram_preset and quality_mode - the two you'll actually change. Start at your VRAM size + balanced.
• prefer_low_ram and allow_spatial_tiling - leave both on unless you have headroom to spare.

Install

Part of IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Or ComfyUI Manager → "IAMCCS" → install → restart. This node leans on native LTX/LTX-Video nodes already in ComfyUI, so no model downloads beyond your usual LTX-2 setup.

Gotchas

Don't run the Atelier without checking what the sampler pipeline upstream gave you - feed it a low-res, short, already-decent clip first and confirm the report numbers (tile sizes, overlaps) match your VRAM before going long. "Quality" mode on 8GB is slow; that's expected. And the -1 defaults you'll see on some Atelier knobs mean "inherit from the preset," not "invalid" - leave them until you're confident you know better.