This is the Minimax workflow, an optimization that reduces generation time while maintaining quality.

It is difficult to install directly because it requires several nodes and complex backend dependencies.

Although all documentation exists within the workflow,

it is highly likely to be very difficult if a human attempts to implement it manually.

Please instruct an AI agent, such as Codex or Claude Code, to implement it.

PASTE to AGENTS
-----------------------------

# AI 에이전트에게 붙여넣으세요

아래 English Agent Instruction 블록 전체를 Codex, Claude Code 또는 로컬 PC를 직접 다룰 수 있는 코딩 에이전트에게 붙여넣으세요. 에이전트가 Civitai에서 최신 ZIP을 받고, 사용자의 ComfyUI 환경을 확인한 다음 필요한 노드·커널·모델을 설치하고 실제 5초 영상까지 검증합니다.

워크플로우 페이지:

https://civarchive.com/models/2890123/minimax-speed-x-6-workflow-sagesollatent-upscale

---

## English Agent Instruction

```text

Your mission is to install and verify the latest MiniMax Speed x6 R2V workflow in my existing ComfyUI installation. Work directly on the real installation. Do not stop at a plan or ask me to perform ordinary installation steps. Preserve my unrelated files, workflows, models and running jobs.

SOURCE PACKAGE

Download the newest ZIP from this Civitai page:

https://civarchive.com/models/2890123/minimax-speed-x-6-workflow-sagesollatent-upscale

Use the package named MiniMax_Speed_x6_Workflow_v10_with_Sol_Adapter.zip, or the newest file whose extracted contents match this structure:

minimaxSpeedX6Workflow_v10.json

sol_attn_minimax_pr117/

└── init.py

Expected integrity values for the current v10 package:

- ZIP SHA-256: FD5C08C11F9304E04D4406B57B50F914EF3DBEC692DE743A0003D45DFFDA43F5

- Workflow SHA-256: CA7D4C611B340A7D5584D046585470E5737D51A7331DF7F67273F1AA6231406E

- Sol adapter SHA-256: 2A2D2C99AC6790B2FC5DD7519FE3436B314A32C18B68C05D92794F9C3B78861B

Civitai may rename the downloaded archive. The filename may differ, but do not continue unless the extracted workflow and adapter hashes match. Do not use an older JSON-only download as a complete package.

AUTHORITY AND NON-NEGOTIABLES

1. Extract the ZIP and read the complete Markdown note titled “Build This Workflow — Codex / Claude” inside the workflow before installing. That embedded note is the detailed implementation authority. This prompt supplies the high-level execution order.

2. Keep the shipped generation contract unchanged: 20 total steps, simple scheduler, res_multistep sampler, split 15+5, 3D latent upscale x2 and joint video/audio decode.

3. Do not add Fused Modulation, Turbo LoRA, Spectrum, Sigma Shift or Ultimate SD Upscale.

4. Do not substitute Saganaki ComfyUI-sol-attn, ComfyUI-SolAttn_triton/SolAttnPatch, Spectrum or another attention implementation for the required KJ PR #117 kernel and bundled DENO adapter.

5. Do not rename PathchSageAttentionKJ. The misspelling is the real serialized KJNodes node ID.

6. Do not create a separate ComfyUI test installation, backup package or migration dossier. Use the existing installation and make ordinary reversible installation decisions yourself. Ask only if a missing NVIDIA driver, CUDA Toolkit or system compiler requires a system-level installation/change, or if an actual destructive conflict is discovered.

EXECUTION ORDER

1. Detect the real ComfyUI root and the exact Python executable used by it. Identify Desktop, portable or Git/venv layout. Record OS/CPU architecture, Python ABI, Torch version, torch.version.cuda, NVIDIA driver, GPU name, VRAM and torch.cuda.get_device_capability(). Check the current queue before restarting ComfyUI.

2. Update ComfyUI first using the supported method for that installation and finish its core requirements installation before installing Sol. Use the exact ComfyUI Python for every pip, import, build and smoke command. Never assume a fixed python_embeded path.

3. Install or update these required node packs under ComfyUI/custom_nodes, using ComfyUI Manager when it resolves them or their official repositories otherwise. Install each repository’s requirements with the exact ComfyUI Python:

- https://github.com/Deno2026/comfyui-deno-custom-nodes

- https://github.com/kijai/ComfyUI-KJNodes

- https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

- https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

ComfyUI-VFI and the bypassed RTX finishing island are optional and must not block the main workflow.

4. Copy the bundled workflow to ComfyUI/user/default/workflows/minimaxSpeedX6Workflow_v10.json, or load it directly from the extracted package.

5. Copy the complete bundled adapter folder to exactly:

ComfyUI/custom_nodes/sol_attn_minimax_pr117/

The resulting file must be:

ComfyUI/custom_nodes/sol_attn_minimax_pr117/__init__.py

Verify its SHA-256 is 2A2D2C99AC6790B2FC5DD7519FE3436B314A32C18B68C05D92794F9C3B78861B. Do not regenerate or replace this adapter with KJ’s temporary sol_attn_minimax_v3.py; that upstream helper has a different saved interface and lacks this package’s long-context protection.

6. Install SageAttention separately for the detected Python, Torch, CUDA and GPU environment. KJNodes provides PathchSageAttentionKJ but does not guarantee a compatible SageAttention package. Install matching Triton/Windows Triton when required. Verify import sageattention with the exact ComfyUI Python. Preserve the workflow values sage_attention=auto and allow_compile=false.

7. Install Kijai’s comfy-kitchen Sol kernel from Comfy-Org/comfy-kitchen PR #117 after all ComfyUI/core requirements. The tested source is branch kijai:sol_attn at commit:

23d1a665ecafd8a216a50108207816d1e133d646

PR: https://github.com/Comfy-Org/comfy-kitchen/pull/117

Source branch: https://github.com/kijai/comfy-kitchen/tree/sol_attn

Use a PR CI wheel only when its OS/CPU, Python ABI, Torch/CUDA runtime, NVIDIA driver and compiled GPU architecture match this PC. Install it with --no-deps --force-reinstall because the PR build can share its version number with stock comfy-kitchen. Never install an sm_120-only wheel on another GPU.

If no matching wheel exists, build the pinned source with recursive submodules and COMFY_CUDA_ARCHS set from the actual torch.cuda.get_device_capability(). Source building requires NVIDIA compute capability 8.0 or newer, BF16 support, CUDA Toolkit 12.8+ with nvcc, CMake 3.26+, Ninja and a supported host compiler/Windows SDK. A minimal build-system compatibility fix may be used and reported, but do not delete Sol sources, add fake/stub symbols or report eager/SDPA/Sage fallback as a successful Sol build.

8. Verify the native kernel with the exact ComfyUI Python. Package version alone is not evidence. The following direct CUDA backend test must pass:

import torch

import comfy_kitchen as ck

from comfy_kitchen.backends import cuda as ck_cuda

assert callable(ck.sol_attn)

assert callable(ck_cuda.sol_attn)

assert callable(ck_cuda.sol_attn_workspace_bytes)

q, k, v = [torch.randn((1, 256, 1, 128), device="cuda", dtype=torch.bfloat16) for _ in range(3)]

out = ck_cuda.sol_attn(q, k, v, tau=1.3)

torch.cuda.synchronize()

assert out.shape == q.shape and out.dtype == torch.bfloat16 and torch.isfinite(out).all()

Do not test using import comfy_kitchen.sol_attn; sol_attn is a callable exposed by the package, not a Python submodule.

9. Download the five required MiniMax H3 files from the workflow’s “Model Links” note and place them in the exact folders listed there: two VAEs, the ref2va diffusion model, one Qwen3-VL text encoder and minimax_h3_latent_upscaler_3d_bf16.safetensors. Make every loader resolve to a file that exists.

10. Restart ComfyUI after Python wheels, custom-node Python or model discovery changes. Verify these main-path node IDs through /object_info or the startup log:

DenoMiniMaxH3ReferenceToVideo

DenoMiniMaxH3ReferenceImageLoader

DenoResolutionSetup

DenoTextEncoderUnload

PathchSageAttentionKJ

SolAttnMiniMax

MinimaxH3LatentUpscaler3D

LTXVSeparateAVLatent

LTXVConcatAVLatent

VHS_VideoCombine

Confirm the model chain is UNETLoader -> PathchSageAttentionKJ -> SolAttnMiniMax. Confirm SolAttnMiniMax comes from custom_nodes.sol_attn_minimax_pr117 and exposes the exact schema saved in the embedded build note.

11. Run one real smoke test. Use one valid local reference image and a short six-section H3 R2V prompt. Keep optional finishing islands bypassed. For this first run only, set Sol verbose=true and use the shipped preset:

- 0.15 MP, 16:9

- 512×288 start

- 1024×576 final after x2 latent upscale

- 5 seconds

- 20 steps split 15+5

- ref_image_size=match

Require at least one “[sol_attn] sparse ... cuda-int8” log line, no “[sol_attn] kernel failed”, no CUDA illegal-memory error and a playable MP4 containing both video and audio. Queue acceptance alone is not completion. Restore verbose=false afterward.

12. Finish with a concise report containing the ComfyUI root, exact Python/Torch/CUDA/GPU, KJ PR commit and wheel source/build, adapter hash, selected text encoder and final MP4 path. If the GPU is unsupported or the direct CUDA test fails, report the exact blocking condition instead of silently substituting another implementation.

```

---

## 한국어 에이전트 지시서

```text

내 기존 ComfyUI에 최신 MiniMax Speed x6 R2V 워크플로우를 직접 설치하고 실제 출력까지 검증하라. 계획만 설명하거나 일반적인 설치 작업을 사용자에게 떠넘기지 말고, 기존의 관련 없는 파일·워크플로우·모델·실행 중인 작업은 보존하라.

다음 Civitai 페이지에서 최신 ZIP을 다운로드하라:

https://civarchive.com/models/2890123/minimax-speed-x-6-workflow-sagesollatent-upscale

사용할 파일은 MiniMax_Speed_x6_Workflow_v10_with_Sol_Adapter.zip이며, 압축을 풀었을 때 반드시 다음 구조여야 한다:

minimaxSpeedX6Workflow_v10.json

sol_attn_minimax_pr117/

└── init.py

현재 v10 검증값:

- ZIP SHA-256: FD5C08C11F9304E04D4406B57B50F914EF3DBEC692DE743A0003D45DFFDA43F5

- 워크플로우 SHA-256: CA7D4C611B340A7D5584D046585470E5737D51A7331DF7F67273F1AA6231406E

- Sol 어댑터 SHA-256: 2A2D2C99AC6790B2FC5DD7519FE3436B314A32C18B68C05D92794F9C3B78861B

Civitai가 ZIP 이름을 바꾸더라도 압축 내부 파일과 해시가 일치하면 된다. 구형 JSON 단독 파일을 완전한 패키지로 취급하지 마라.

1. 압축을 풀고 워크플로우 내부의 “Build This Workflow — Codex / Claude” 마크다운 노트를 처음부터 끝까지 읽어라. 그 노트가 상세 구현 정본이며, 이 지시서는 전체 진행 순서를 정한다.

2. 실제 ComfyUI 루트와 그 설치본이 사용하는 정확한 Python을 찾고 OS, Python ABI, Torch, torch.version.cuda, NVIDIA 드라이버, GPU, VRAM, Compute Capability를 확인하라. 재시작 전에는 현재 큐도 확인하라.

3. ComfyUI와 core requirements를 먼저 최신 상태로 업데이트하라. 이후 모든 pip·import·빌드·테스트는 정확한 ComfyUI Python으로 실행하라.

4. DENO Custom Nodes, KJNodes, VideoHelperSuite, MiniMax H3 Latent Upscaler를 설치 또는 업데이트하고 각각의 requirements를 같은 Python에 설치하라.

5. 워크플로우 JSON은 ComfyUI/user/default/workflows/에 넣거나 압축을 푼 위치에서 직접 불러와라.

6. ZIP에 동봉된 sol_attn_minimax_pr117 폴더 전체를 ComfyUI/custom_nodes/ 아래에 복사하라. 최종 경로는 ComfyUI/custom_nodes/sol_attn_minimax_pr117/__init__.py이며 SHA-256은 2A2D2C99AC6790B2FC5DD7519FE3436B314A32C18B68C05D92794F9C3B78861B이어야 한다. KJ의 임시 sol_attn_minimax_v3.py로 교체하거나 어댑터를 다시 만들지 마라.

7. 현재 Python·Torch·CUDA·GPU에 맞는 SageAttention과 필요한 Triton을 별도로 설치하고 import sageattention을 확인하라. PathchSageAttentionKJ의 저장값 auto, allow_compile=false를 유지하라.

8. ComfyUI 업데이트와 requirements 설치가 끝난 다음 KJ comfy-kitchen PR #117의 sol_attn 브랜치 커밋 23d1a665ecafd8a216a50108207816d1e133d646을 마지막에 설치하라. 맞는 CI wheel이 있으면 --no-deps --force-reinstall로 설치하고, 없으면 실제 GPU Compute Capability에 맞는 COMFY_CUDA_ARCHS로 pinned source와 recursive submodule을 빌드하라. 다른 GPU용 wheel을 재사용하지 마라.

9. 정확한 ComfyUI Python에서 comfy_kitchen.sol_attn, CUDA backend sol_attn, sol_attn_workspace_bytes가 callable인지 확인하고 워크플로우 노트에 있는 BF16 직접 CUDA 테스트를 실행하라. stock comfy-kitchen 버전 번호나 import 성공만으로 완료 처리하지 마라.

10. 워크플로우의 Model Links 노트에서 필수 모델 5개를 다운로드해 표시된 폴더에 넣고 모든 loader가 실제 파일을 가리키게 하라.

11. ComfyUI를 재시작하고 /object_info 또는 시작 로그에서 필요한 노드를 확인하라. 모델 연결은 UNETLoader -> PathchSageAttentionKJ -> SolAttnMiniMax이어야 한다. PathchSageAttentionKJ의 오타는 실제 공개 ID이므로 고치지 마라.

12. Fused Modulation, Turbo LoRA, Spectrum, Sigma Shift, Ultimate SD Upscale 또는 다른 Sol/Spectrum 노드를 추가하거나 대체하지 마라. 20 steps, simple, res_multistep, 15+5, 3D latent x2, 영상·오디오 공동 decode 구조를 유지하라.

13. 유효한 로컬 레퍼런스 이미지 하나와 짧은 6섹션 H3 R2V 프롬프트를 준비하고 0.15MP, 16:9, 시작 512×288, 최종 1024×576, 5초, 15+5, ref_image_size=match로 스모크 테스트하라. 첫 실행에서만 Sol verbose=true로 두고 “[sol_attn] sparse ... cuda-int8” 실행 로그, 커널 오류 없음, playable 영상+오디오 MP4를 확인한 다음 verbose=false로 복구하라.

14. 완료 시 ComfyUI 경로, Python/Torch/CUDA/GPU, PR 커밋과 wheel 출처 또는 빌드 내용, 어댑터 해시, 선택한 텍스트 인코더와 최종 MP4 경로를 짧게 보고하라. 지원되지 않는 GPU나 CUDA 테스트 실패가 발생하면 다른 구현으로 숨겨서 대체하지 말고 정확한 중단 이유를 보고하라.

별도 테스트 ComfyUI, 백업 패키지, 장황한 승격 보고서는 만들지 마라. 다만 NVIDIA 드라이버·CUDA Toolkit·시스템 컴파일러 같은 시스템 수준 설치가 새로 필요하거나 실제 파괴적 충돌이 발견되면 그때만 사용자 승인을 받아라.

```