Wan 2.2 Seed and Guidance Phases Explained: What Changes Output and What Doesn't (2026)

You set the same prompt, the same seed, the same everything — and got a completely different result.

Or you changed the seed to try a variation, and the output barely shifted.

Or you opened the ComfyUI Wan 2.2 sampler, saw "guidance" with two separate fields labeled high noise and low noise, and had no idea which one to touch.

Every parameter in Wan 2.2 interacts with every other parameter, but seed and guidance are the two that cause the most confusion — because they are the least documented.

I tested seed behavior across dozens of generations with fixed and random values, documented the GitHub issue where seeds silently fail, and mapped how the two-stage high noise / low noise guidance architecture actually changes output at each phase. This guide covers what seed controls (and what it does not), what "guidance phases" means in Wan 2.2's specific two-model architecture, and the practical strategies that save the most rerolls.

The Short Answer

Seed changes the starting point. Guidance changes how closely the model follows your instructions. They control different things, and you need both to get consistent, usable output.

What a Seed Actually Does (and Does Not Control)

A seed is a number that initializes the random noise pattern the model starts from. Think of it as the "starting canvas" — the same seed always produces the same static noise pattern, and the model denoises that specific pattern into a video.

When you set a seed, you lock this starting canvas. With the same prompt, resolution, frame count, sampler, scheduler, step count, shift value, and guidance scale — the model follows the same denoising path and produces the same or very similar output.

What seed changes about the output:

• Micro-textures and surface detail arrangement
• Motion path variations (a character turns left vs right at the start)
• Which random "interpretation" of your prompt the model chooses
• Minor composition shifts within the same framing

What seed does NOT change:

• Subject matter (your prompt controls this)
• Overall composition (framing, angle — prompt and I2V reference control this)
• Lighting direction (prompt controls this)
• Color palette (prompt controls this)
• Whether the output is good or bad (seed does not encode quality)

What Does "-1 for Random" Mean?

In the ComfyUI Wan 2.2 sampler node (and most other Wan 2.2 interfaces), a seed value of -1 or leaving the field empty means "assign a random seed for this generation." The system generates a random positive integer behind the scenes and uses that as the seed.

The output shows you which seed was used in the console log or result metadata, so you can copy it and reuse it if you want to reproduce or iterate on that specific result.

Rule of thumb: Use -1 during exploration to see what the model produces. Once you get something close to what you want, copy the seed, set it explicitly, and iterate from there by changing one parameter at a time.

Why Different Seeds Sometimes Give the Same Result

This is the most common seed-related issue in Wan 2.2, and it is a known problem.

There is an active ComfyUI GitHub issue (#9114) titled "WAN 2.2 different seeds give the same result." Users report that in certain T2V workflows, changing the seed on the sampler node produces identical or near-identical output across multiple runs.

There are three known causes:

Cause 1: Custom nodes overriding the seed. Some ComfyUI custom nodes reset or ignore the seed value you set on the sampler. This is especially common with workflow templates downloaded from CivitAI or GitHub that include seed-control nodes that conflict with the main sampler.

Fix: Run ComfyUI with --disable-all-custom-nodes to check if the seed starts working. If it does, re-enable custom nodes one at a time to find the conflicting one.

Cause 2: The seed is not wired to the sampler node. In complex ComfyUI workflows with multiple samplers or control nodes, the seed you set on the main interface may not be connected to the actual sampling operation.

Fix: Check the node graph — make sure the seed input on your sampler node is connected to a seed source (or a fixed value node), not floating unconnected.

Cause 3: The workflow uses a fixed seed node internally. Some workflow authors hardcode a seed value inside the workflow, overriding whatever you type into the UI.

Fix: Search the workflow for any node that outputs a seed value. You will usually find a "Seed" node or "Int" node with a fixed number wired to the sampler. Replace it with your own seed source.

Rule of thumb for seed troubleshooting: If changing the seed produces zero visible change, the seed is not reaching the sampler — look for a custom node or hardcoded value before you look at the model. Hardware and environment issues change output slightly between runs, not identically.

Expert pitfall: Even when the seed is working correctly, the same seed does NOT guarantee bit-identical output across different hardware (NVIDIA vs AMD, different CUDA versions) or different ComfyUI versions. The denoising math is deterministic within the same environment, but floating-point differences across GPUs and libraries introduce minor variations. If pixel-perfect reproducibility matters, run all comparison generations on the same machine.

Fixed Seed vs Random Seed: When to Use Each

Rule of thumb for iteration: Fix the seed first, then tune guidance. Once you find guidance values that work, unlock the seed and generate variations. This order saves the most rerolls because it separates "does the prompt work?" from "does this specific noise pattern look good?"

Seed controls where you start. Guidance controls where you go. Now let's cover the part that is unique to Wan 2.2 — and the most misunderstood parameter in its workflow.

What Guidance Phases Actually Mean

"Guidance phases" in Wan 2.2 refers to the fact that the model uses two separate diffusion models (called "experts" in the MoE architecture) that run in sequence — a high noise model and a low noise model — and each can have its own guidance scale and step count.

Why Two Phases?

Wan 2.2 is a Mixture-of-Experts (MoE) model. Instead of one large model doing all the work, it has specialized sub-models. The high noise expert is trained to handle the early, chaotic denoising steps where the main structure emerges. The low noise expert handles the later steps where fine details are added.

The ComfyUI Wan 2.2 native workflow expresses this as two separate Load Diffusion Model nodes — one for the high noise checkpoint and one for the low noise checkpoint. Each runs for a configurable number of steps with its own guidance scale.

What Each Phase Controls

High noise phase (first ~40–60% of steps):

• Establishes broad composition and subject placement
• Determines motion direction and flow
• Sets scene geometry and spatial arrangement
• Guidance here controls how strongly the prompt dictates these large-scale features

Low noise phase (remaining steps):

• Refines textures and surface details
• Adds fine motion细节
• Smooths transitions and reduces noise
• Guidance here controls how closely details match the prompt

How Steps Split Across Phases

The community default for standard workflows is 20–24 total steps split evenly or slightly favoring the high noise phase. LightX2V distilled models use as few as 2 steps per phase.

How Guidance Scale Affects Output

Guidance scale (also called CFG — Classifier-Free Guidance) controls how strongly the model adheres to your text prompt versus "doing its own thing." A higher value forces the output to match the prompt more closely. A lower value lets the model be more creative or natural.

The math behind it is straightforward: at each denoising step, the model makes two predictions — one conditioned on your prompt and one unconditioned. The final prediction is unconditioned + CFG × (conditioned − unconditioned). At CFG 1.0, the conditioned term cancels out and the model ignores your prompt entirely. At CFG 7.0, the prompt term dominates and any difference between conditioned and unconditioned is amplified, which is why high CFG values produce oversaturated, "plastic" output. The sweet spot at 3.5 means the conditioned prediction is weighted 3.5× more than the gap between conditioned and unconditioned — enough to follow the prompt without exaggerating every difference.

Rule of thumb for guidance: Start at 3.5 for every new prompt. Go up if the prompt is ignored. Go down if the output looks artificial or oversaturated. Change in steps of 0.5 and evaluate one generation at a time.

Expert pitfall for two-phase guidance: The high noise and low noise phases do not have clean boundaries. A guidance change in the high noise phase can alter the latent state in ways that the low noise phase cannot fully correct. If you set high noise guidance aggressively high (5.0+) and then set low noise guidance low (2.0) expecting it to "fix" the oversaturation, it will not — the damage is done in the first phase. Always treat the two phases as a coupled system, not independent levers.

Separate Guidance for Each Phase

In the ComfyUI Wan 2.2 native workflow, you can set different guidance values for the high noise and low noise phases. This lets you control how strictly the prompt is followed during the broad composition phase versus the fine detail phase.

Common patterns:

• Equal guidance (both at 3.5): Standard approach. Works for most prompts.
• Higher high noise, lower low noise (4.0 / 3.0): Forces strong prompt adherence for composition but lets details be more natural. Good for action prompts where you want the motion direction to match the prompt but the textures to look organic.
• Lower high noise, higher low noise (3.0 / 4.0): Allows creative composition but forces details to match. Useful for abstract prompts where the broad scene can be loose but the specific elements (like a character's face) must be precise.
• High noise only (low noise at 1.0): Effectively disables the low noise refinement. Output looks rough and unfinished. Not recommended.

Practical Strategies for Seed + Guidance

Combine seed and guidance deliberately to reduce rerolls and get consistent results.

Strategy 1: Seed-First Iteration

1. Start with a random seed (-1) and default guidance (3.5 for both phases)
2. Generate until you find a seed that produces a composition you like
3. Fix that seed
4. Tune guidance values — high noise first, then low noise
5. Once guidance is dialed in, unlock the seed and generate variations

This is the most efficient workflow because it isolates variables. Changing seed and guidance at the same time means you never know which one caused the change.

Strategy 2: Baseline Clipboard

Maintain a set of "baseline seeds" — 3–5 seeds that you have tested with your default prompt structure and known to produce reliable results. When you try a new prompt variation, test it against all baseline seeds before evaluating quality. If all baseline seeds produce similar degradation, the prompt is the problem. If only one baseline seed degrades, the seed is the problem.

Strategy 3: High-Low Guidance Separation

If you consistently need to adjust output quality:

• Set high noise guidance to control motion and composition strength
• Set low noise guidance to control texture and detail quality
• Adjust high noise first (it has a larger effect on the output)
• Adjust low noise only if details look off after high noise is correct

This separation is unique to Wan 2.2 and does not exist in single-model video generators. Using it deliberately gives you finer control than any single CFG slider can provide.

Expert pitfall for all strategies: The baseline clipboard strategy breaks if you change your sampler or scheduler after establishing baseline seeds. Seeds are only reproducible within the same sampler/scheduler/shift combination. A seed that produces a great composition with Euler + Simple will produce completely different motion with UniPC + Simple — the seed still determines the starting noise, but the sampling trajectory diverges immediately. If you change your sampler, re-establish your baseline seeds.

What Not to Do: Common Mistakes That Waste Generations

After you understand what seed and guidance control, the most important thing is knowing what NOT to do.

Do not chase the "perfect seed." There is no universal good seed. A seed that produces a great composition with one prompt can produce garbage with a slightly different prompt. Spending hours testing seeds before writing your prompt is backward — write the prompt, find what works, then lock the seed.

Do not change seed and guidance at the same time. Every reroll that changes both parameters tells you nothing about which one caused the difference. Change one variable per generation. The single biggest source of wasted credits is changing everything at once and hoping for a different result.

Do not set guidance above 5.0 as a default. High guidance (6.0+) produces visible artifacts — oversaturation, flickering, "plastic" motion. It is not a "better" setting. It is a specialized tool for prompts that consistently underperform at normal values. If your output looks bad at 3.5, fix the prompt before you raise guidance.

Do not assume your seed is being used just because you set it. As covered in the troubleshooting section above, custom nodes, workflow templates, and unconnected seed inputs can silently override your seed. Always verify by generating twice with the same seed — if the outputs are not near-identical, something is overriding your seed.

Do not reuse baseline seeds across different samplers or schedulers. The relationship between seed and output is only stable within the same sampler + scheduler combination. Change either one, and your baseline seeds are meaningless.

Rule of thumb for saving rerolls: Every generation should change exactly one variable. If you change the prompt, keep the seed and guidance the same. If you change guidance, keep the seed and prompt the same. If you change the seed, keep everything else the same. Three generations per variable, then evaluate.

Frequently Asked Questions

What is the best seed for Wan 2.2? There is no best seed. Seeds do not encode quality — they determine the starting noise pattern. A seed that produces a great result with one prompt may produce a poor result with another. The same seed with the same prompt and parameters will produce the same result reliably.

Does seed work differently in Wan 2.2 than in Stable Diffusion? The core mechanism is the same — seed initializes noise. The difference is that Wan 2.2's two-stage (high noise + low noise) architecture means the seed interacts with two separate denoising paths, which can make the output more sensitive to step counts and guidance values per phase. In practice, seed behavior is consistent between the two, but the interaction with guidance phases is specific to Wan 2.2.

Can I use a negative seed value? In most ComfyUI Wan 2.2 implementations, negative values are treated as -1 = random. Setting -2 or any other negative number usually resolves to the same random behavior as -1. Use only positive integers or -1.

What sampler and scheduler work best with guidance phases? The community standard for Wan 2.2 is UniPC sampler with Simple scheduler, or Euler sampler with Simple scheduler. The LightX2V distilled models work best with LCM sampler and ddim_uniform scheduler. Sampler choice interacts with guidance — UniPC is more sensitive to guidance values than Euler, so if you change your sampler, retune your guidance.

Does guidance scale affect generation speed? No. Guidance scale changes the math inside each step but does not change the number of steps or the inference time. Only step count and resolution meaningfully affect generation speed.

How do I check which seed was used after a random generation? In ComfyUI, the console log shows the seed used for each generation. In the Wan 2.2 native workflow, the sampler node outputs the seed value after generation — you can connect a display node to see it in the UI. If neither option works, search your ComfyUI log for "seed" to find the value.

Why does my output change when I change only the guidance value? Guidance changes how the model interprets the prompt at every step. Even with the same seed, a different guidance value produces a different denoising path from step 1 onward. The seed controls the starting noise, but guidance controls every decision after that. You cannot change guidance and keep everything else the same — the output will always differ.

Does the fixed seed guarantee identical output across two different computers? No. Differences in GPU architecture, CUDA version, PyTorch version, and floating-point math can introduce small variations that compound across steps. For truly reproducible results, run on the same hardware and software environment.

Summary

Most Wan 2.2 users waste generations changing everything at once and learning nothing. Seed and guidance are separate controls, and understanding the split eliminates most of that waste.

• Seed (-1 for random): Controls the starting noise pattern. Same seed + same params = same or very similar output. Use fixed seeds for iteration, random seeds for exploration.
• Guidance scale (CFG): Controls how strongly the model follows your prompt. 3.5 is the default sweet spot. Go up for stricter adherence, down for more natural output.
• Guidance phases: Wan 2.2 runs two models in sequence — high noise (broad composition) and low noise (fine details). Each phase can have its own guidance value, giving you finer control than a single CFG slider.

The most common mistake is changing seed and guidance at the same time and trying to judge the result — you cannot isolate what changed. Fix the seed first, tune guidance per phase, then unlock the seed for variation.

Next step: If you are new to Wan 2.2 parameters, start with the Wan 2.2 Prompt Guide to understand how prompt structure interacts with seed and guidance. For the full parameter reference including step count, sampler, and scheduler recommendations, the Wan 2.2 ComfyUI Workflow Guide covers the complete setup.