KSampler Progress (Inspire)

You know that satisfying clip where a noisy blur resolves step by step into a finished image? This is the node that makes it. KSampler Progress is a regular KSampler that, on the side, captures the latent at intervals through the denoising process and outputs the whole sequence as a batch. Pipe that into a Video Combine node and you've got a timelapse of your generation forming. It's a favorite for social posts and for genuinely seeing what your sampler and scheduler are doing across steps.

Beyond looking cool, it's a real diagnostic. Watching where an image locks in - early, late, oscillating - tells you a lot about whether your step count is wasted or your scheduler is fighting you. That's why it lives in the pack's "analysis" category, not just its toy drawer.

How it works

It samples normally to produce your final latent, but every interval steps it also grabs the intermediate latent and accumulates them. The final output is two things: your finished image's latent, and a batch of the progress latents. Decode that batch through a VAE and each frame is one snapshot of the denoising; string them together and it's motion. The README points you at Video Combine from ComfyUI-VideoHelperSuite to turn the batch into an actual video file.

The inputs and outputs that matter

Most of it is a standard KSampler - model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise - set those as you would anywhere. The ones specific to this node:

• interval (default 1) - capture a progress frame every N steps. At 1 you get every single step (smoothest video, most frames); raise it to thin the sequence for a shorter clip.
• omit_start_latent (default on) and omit_final_latent (default off) - whether to drop the pure-noise first frame and the finished last frame from the progress batch. Defaults are sensible: skip the boring all-noise opener, keep the finished frame.
• noise_mode (GPU(=A1111) / CPU / and internal-seed variants) - where the initial noise is generated. This is Inspire's A1111-reproduction feature: A1111 makes noise on the GPU, stock ComfyUI on the CPU, and matching it matters if you're chasing a specific seed's result.

Two outputs: latent (your actual result - wire to VAE Decode) and progress_latent (the batch - wire to a VAE Decode then Video Combine).

Installing it

Part of the Inspire Pack by Dr.Lt.Data, who wrote ComfyUI Manager and the Impact Pack - mainstream, well-kept tooling. Via ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

then restart. To actually export a video you'll also want ComfyUI-VideoHelperSuite for its Video Combine node - install that via Manager too.

Common issues

The progress batch eats memory. Every captured step is a full latent held in memory, and at interval 1 with high step counts that's a lot of frames. If you OOM or things crawl, raise interval so it captures fewer snapshots.

No video comes out. This node only produces latents. You have to VAE-decode the progress_latent batch and run it through Video Combine (or Save Image as an animated batch) - the node itself doesn't write a file.

My result doesn't match a bare KSampler. Check noise_mode. If your reference was made in A1111, use GPU(=A1111); if you want to match stock ComfyUI, use CPU. Seeds don't reproduce across noise sources.

Nodes won't load after install. Read the terminal log - usually a version mismatch. Update Inspire Pack and Impact Pack together and restart.