ComfyUI-Sol-Attn Visual tour

The annotated screenshots show real nodes in ComfyUI 0.38.0 with frontend 1.53.10, Nodes 2.0, and Sol-Attn v0.6.7. Values are defaults except the explicitly marked dense_blocks and dense_percent examples. The Visual tour is also available at GitHub.

MiniMax H3 Workflow

Sol-Attn sparse attention

Search: Sol-Attn (sparse attention) · Category: model_patches/attention

Connect loader → Sol-Attn → guider/sampler. This generic node patches ComfyUI's attention hook on the connected model branch. Unsupported shapes, masks, or sequences below min_tokens use the captured fallback backend.

• tau controls approximation: higher values send more attention blocks through the approximate path. Start from the default 1.3 and compare the same inputs and seed.

• min_tokens defaults to 4096. strict = true raises kernel errors for diagnosis instead of taking the error fallback.

• thresh_type = diag is the default estimator; exact uses more detailed statistics at additional cost.

• Both INT8 switches default to off. Enable int8_qk before int8_pv; their quantization error is additional to the sparse approximation.

For an H3 workflow using a module-level memory-efficient Sage patch, choose one of the following H3 Sol nodes. That Sage patch replaces the attention forward and bypasses the hook used by this generic node.

MiniMax H3 Memory Efficient Sol Attention Patch

Category: model_patches/attention

This H3-specific node passes views of the fused QKV projection directly to Sol. tau stays constant during sampling. Use this node or Scheduled Sol Attention on a model branch.

• Keep sink_conditioning = exact_kv to preserve exact conditioning KV blocks. exact_kv_and_rows also keeps their query rows dense; off disables this protection.

• dense_blocks uses 0-based indices. The illustrated 0-2,-1 keeps the first three and final blocks on the fallback; its actual default is empty.

• Optional H3 Sage patches belong before the H3 Sol patch, so Sol can capture their forward as its fallback. A Sage patch applied afterward shadows Sol.

The short-sequence gate, estimator, strict mode, and INT8 controls serve the same purposes as in the generic node.

MiniMax H3 Scheduled Sol Attention Patch

Category: model_patches/attention

Use the same MODEL placement as fixed-tau H3 attention. Defaults move tau from 1.3 at high noise to 0.8 near the end, with a linear curve. Lower tau retains more exact attention. Setting both tau values equal gives a constant threshold.

• curve selects how tau moves between the endpoints: linear, cosine, sqrt, smoothstep, exponential, or step.

• dense_percent keeps the captured fallback for an initial fraction of sampling. The screenshot demonstrates 0.20; the default is 0.0.

• Connect tau_graph → Preview Image to inspect the schedule. Send the model output to the next patch or guider.

• Conditioning sinks and dense-block exclusions remain available. The shown 0-2,-1 is an example, not a default.

MiniMax H3 Fused Modulation

Category: model_patches/optimization

Connect this after your H3 attention patch and before Chunk FeedForward for a readable chain. It fuses modulation and gated residual operations while reproducing the intermediate BF16 rounding of the supported eager path.

The only control is enabled. It does not select an attention backend: installed attention and feed-forward patches remain in use. Unknown whole-block forward patches are left untouched. Switch it off to compare the same workflow without fusion.

MiniMax H3 Chunk FeedForward

Category: model_patches/memory

This node splits token-local feed-forward work into chunks and keeps the model's existing linear-input activation implementation. It can be used independently of Sol attention.

• chunks = 2 is the default. More chunks lower the intermediate memory peak but add overhead.

• min_tokens = 8192 keeps shorter sequences on the normal MLP. Lower it if you need chunking for shorter renders.

• Memory savings depend on the checkpoint. The INT8 ConvRot path benefits most; plain BF16 retains additional intermediates and sees smaller savings.

Send the MODEL output onward to optional conditioning or cache patches, then the guider or sampler.

MiniMax H3 Diff-Aid Text Modulation

Search: MiniMax H3 Diff-Aid Text Modulation (Experimental) · Category: model_patches/conditioning

Place this after the other H3 patches. It modifies language rows in place, using text-sized scratch memory, so the Comfy compiler can remain enabled. Replace the external H3 Diff-Aid node in this branch; leaving that node connected retains its full-stream clone and AIMDO failure.

• block_indices uses 1-based indices, unlike attention's dense_blocks. The default 1,13,25,37,50 is an experimental selection.

• strength = 0.20 scales the selected language features. Zero strength or enabled = false bypasses the patch.

• sigma_start and sigma_end bound the active noise window: 0 is no noise, 1 is this run's highest sigma. Defaults span the whole run. A custom window requires the sampler's sample_sigmas metadata; sigma_ramp adds smooth shoulders outside the window.

• token_weight_mode = none applies uniform strength. linear and exponential taper language-token weights toward token_tail, without modifying visual, reference, audio, or target-video rows.

• cond_only = true changes only a confirmed conditional branch. Unconditional, unknown, or mixed branches bypass modulation.

Use Show advanced inputs to reveal sigma_ramp, token_tail, and cond_only as shown. This inference heuristic has no trained Aid weights or established H3 quality benefit; compare enabled and disabled results with identical inputs and seed.