🌀 Echo DMD Sigmas

Every distilled LTX-2.3 workflow needs a sigma schedule, and DMD models are pickier than most about what you give them. EchoDMDSigmas is the pack's way of handing you the exact schedules the JoyAI-Echo DMD checkpoints were trained around - with a couple of refinement variants on top. If you've been typing sigma lists into a ManualSigmas node by hand and eyeballing them, this node replaces that whole exercise.

The headline fact: Echo's official schedule is a 9-value, two-zone structure. Five tight init steps cluster sigmas from 1.0 down to 0.975 (the "init cluster," where the model sorts out gross structure cheaply), then four large jumps land on the DMD anchor points - 0.909375, 0.725, 0.421875, 0.0. Those anchors are where the model was actually trained to denoise, and they're the values that matter.

The presets

• official - the exact 9-sigma schedule Echo ships with. Start here.
• bridge / minimal - drop an init step and add a midpoint at 0.817 between the two biggest productive gaps. Minor variations on official.
• 10step, 11step, 12step - extended refinement schedules that insert substeps into the two big low-sigma gaps (0.725→0.422 and 0.422→0.0). That's where fine-motion ghosting and softness originate. All DMD anchor points are preserved in every variant, so the model still sees its trained values.

The rule of thumb the author spells out in the source: try 10step first if you see ghosting or mushiness; escalate to 11step/12step only if it doesn't clear. Every added step preserves the anchors, which is the whole trick - you're refining between trained points, not inventing new ones.

There's also custom, which lets you type your own comma-separated list into custom_sigmas. The default string shows the official schedule, so you can see the format. If you go custom, keep the anchors (0.975, 0.909375, 0.725, 0.421875, 0.0) in place and insert around them - the model's adaln timestep embedder only truly knows those values.

Wiring it

Output is a single sigmas SIGMAS tensor. Feed it to both the guider and the sampler in a SamplerCustom setup:

EchoDMDSigmas → (sigmas) → LTX2STGGuider and SamplerCustom
EchoDMDSampler → (sampler) → SamplerCustom

And here's the pairing that matters: if you use anything other than official, run the output through EchoDMDSigmaRemap (interpolate mode) before it reaches the guider and sampler. The remap node re-expresses every sigma in terms of the DMD anchors for model conditioning, while the euler step still uses the real value. Skip it and the extended presets silently feed the model timesteps it never saw - you'll get drift, not refinement. official passes through cleanly without remapping.

Install is the pack-wide one-liner (git clone into custom_nodes as 10S_Nodes, restart, or use ComfyUI Manager and search "10S-Comfy-nodes"). No extra dependencies. And like every node in this pack, it's LTX2-specific - the anchors and two-zone structure only make sense for the JoyAI-Echo DMD family.