# I've hit a wall - please help

**URL:** <https://discuss.huggingface.co/t/ive-hit-a-wall-please-help/175450>\
**Category:** Beginners\
**Created:** [April 21, 2026, 5:58pm UTC](https://discuss.huggingface.co/t/ive-hit-a-wall-please-help/175450 "2026-04-21T17:58:41Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![kdalts3759](https://avatars.discourse-cdn.com/v4/letter/k/b782af/32.png) [@kdalts3759](https://discuss.huggingface.co/u/kdalts3759)\
**Post date:** [April 21, 2026, 5:58pm UTC](https://discuss.huggingface.co/t/ive-hit-a-wall-please-help/175450/1 "2026-04-21T17:58:41Z")

</div>

So I’m now about 3 weeks in to my ComyUi experience.. and I love it.. but I feel like I’ve hit a wall.

- struggling to find workflows that work
- downloaded tons of safetensors, loras, checkpoints that has gobbled up my disk space!
- created 2 min long stitched together videos that look great but the characters at the start of the video look entirely different by the end of the video

now I’ve discovered that some basics are missing from my build ie: ComfyUI manager and found that none of the essentials (from the essentials button) are enabled. I hit install but they don’t enable. In fact, comyi manager status states it is installed and the directory exists in the custom\_nodes folder but there is no button anywhere to launch it…

- I tried installing other essential… goes through the download, get the ‘installed’ message… but not installed
- tried re-installing comfyUI… no change

Not sure how to proceed so any advice (in simple beginner language (remember when you first started!!)) would be most welcome

---

<div class="post-metadata">

**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\
**Post date:** [April 22, 2026, 2:08am UTC](https://discuss.huggingface.co/t/ive-hit-a-wall-please-help/175450/2 "2026-04-22T02:08:27Z")

</div>

Continued from [1st movie clip! - #7 by kdalts3759](https://discuss.huggingface.co/t/1st-movie-clip/175306/7). I haven’t used ComfyUI myself, so I can’t say for sure, but…:

* * *

You can make this work **without** a better GPU right now.

What is happening is that you have hit **two different walls at the same time** :

1. a **tooling wall**  
ComfyUI Manager / ReActor / custom-node install problems

2. a **workflow wall**  
long stitched video + identity drift across clips

Those are separate problems, and they need to be solved in the right order.

* * *

# The most important thing first

## Reinstalling ComfyUI usually does **not** fix this kind of problem

Why not?

Because the usual causes are:

- wrong folder
- wrong Python environment
- missing node dependencies
- frontend cache / missing Manager UI
- a custom node that downloaded but did not actually import

The official ComfyUI custom-node docs explicitly separate **installing the node files** from **installing dependencies** , and they tell you to restart ComfyUI, refresh the browser, and check startup logs for failed imports. ([docs.comfy.org](https://docs.comfy.org/development/core-concepts/custom-nodes))

So the fact that reinstalling ComfyUI changed nothing actually fits the normal failure pattern.

* * *

# Your biggest clue

You said ReActor is in the **`custom_model`** folder.

That is almost certainly the wrong place.

For ReActor, the official install path is under:

```plaintext
ComfyUI/custom_nodes/ComfyUI-ReActor

```

not `custom_model`, not `models`, not a safetensors folder. The ReActor README documents manual installation as a **custom node repo** inside `custom_nodes`, followed by a requirements install step. ([github.com](https://github.com/Gourieff/ComfyUI-ReActor))

So the first big correction is:

- **custom\_models / models** = where model files go
- **custom\_nodes** = where node repositories go

If a node repo is in the wrong place, ComfyUI will never load it as a node.

* * *

# What I think is happening in your setup

In simple language:

## Problem A — Manager is probably on disk, but its UI is not loading properly

That would explain:

- “installed”
- folder exists in `custom_nodes`
- but **no button anywhere**

There are similar Manager issues where the button disappears after updates or frontend/menu changes, even though the repo is present. ([github.com](https://github.com/Comfy-Org/ComfyUI-Manager/issues/1035))

## Problem B — ReActor is either in the wrong folder, or its Python packages were never installed correctly

That would explain:

- downloaded from GitHub
- seems present
- but no node in search
- reinstalling ComfyUI does nothing

There are multiple ReActor reports online where users say it is “installed” but still fails to show up because the import fails or the install script never completed properly. ([github.com](https://github.com/Gourieff/ComfyUI-ReActor/issues/216))

## Problem C — your workflow is asking too much of long stitched clips

That would explain:

- good-looking short clips
- but characters drift by the end of the stitched sequence

For that part, the fix is not “download more LoRAs.” The fix is **anchor-based continuity** : I2V, first/last-frame control, and repair of bridge frames. The official Wan tutorials are already built around those kinds of controls. ([docs.comfy.org](https://docs.comfy.org/tutorials/video/wan/wan-flf))

* * *

# What I would do, in exact order

Do **not** try to fix everything at once.

Use this order.

* * *

## Step 1 — stop adding more models for now

Right now, more downloads are hurting you.

They are:

- eating disk
- adding confusion
- making it harder to tell what works

For the next few days, keep only a **small working set** :

- one base workflow
- one base model
- one or zero LoRAs
- three anchor stills

That is enough.

* * *

## Step 2 — fix the folder structure first

Your `custom_nodes` folder should look roughly like this:

```plaintext
ComfyUI/
  custom_nodes/
    ComfyUI-Manager/
    ComfyUI-ReActor/

```

It should **not** look like this:

```plaintext
ComfyUI/
  custom_model/
    ComfyUI-ReActor/

```

or this:

```plaintext
ComfyUI/
  custom_nodes/
    ComfyUI-ReActor-main/

```

or this:

```plaintext
ComfyUI/
  custom_nodes/
    ComfyUI-ReActor-main/
      ComfyUI-ReActor/

```

That last “nested GitHub zip folder” mistake is extremely common.

* * *

## Step 3 — fix ReActor the official way

After moving ReActor into `custom_nodes`, install its requirements using the **same Python that ComfyUI itself uses**.

That part matters a lot.

The ReActor README shows manual install using ComfyUI’s Python and its `requirements.txt`, and the repo also provides helper scripts such as `install.bat`. ([github.com](https://github.com/Gourieff/ComfyUI-ReActor))

### If you use ComfyUI portable on Windows

Use the Python inside your ComfyUI folder, not a random Python on your PC.

Something like:

```bat
python_embeded\python.exe -s -m pip install -r ComfyUI\custom_nodes\ComfyUI-ReActor\requirements.txt

```

or go into the ReActor folder and run:

```bat
install.bat

```

The exact wording comes from the project’s own install instructions. ([github.com](https://github.com/Gourieff/ComfyUI-ReActor))

### Why this matters

A very common beginner problem is:

- repo is in the right place
- but requirements were installed into the **wrong Python**
- so the node still fails import when ComfyUI starts

That exact pattern comes up again and again in similar cases online. ([reddit.com](https://www.reddit.com/r/comfyui/comments/1and6s9/installing_nodes_but_without_the_manager/))

* * *

## Step 4 — launch ComfyUI from a terminal and read the startup text

This is the part most beginners skip.

The startup console is the truth.

Look for anything like:

- `IMPORT FAILED`
- `ModuleNotFoundError`
- `No module named ...`
- `ComfyUI-Manager`
- `ReActor`

If a custom node downloaded but did not import, the console usually tells you why.

The official docs tell you to check startup logs for failed imports, and real issue reports show that a folder being present means nothing if startup still says `IMPORT FAILED`. ([docs.comfy.org](https://docs.comfy.org/development/core-concepts/custom-nodes))

* * *

## Step 5 — hard refresh the browser

Do this after restarting ComfyUI:

- **Ctrl + F5** on Windows

Why?

Because Manager has a UI/frontend part. Sometimes the backend is there, but the browser is still showing an older cached interface.

This is especially worth trying if:

- the folder exists
- startup looks mostly okay
- but the Manager button still never appears

* * *

# If Manager is still missing

If Manager still refuses to show its button, do **not** let that stop the whole project.

You can still work manually.

That is important.

Manager is helpful, but it is not required for the workflow itself.

So if Manager keeps failing, the priority becomes:

1. get one custom node working manually
2. get the video workflow working
3. come back to Manager later

* * *

# How to make the workflow work **without** ReActor

This is the part I really want to stress:

## ReActor is optional for your project

Useful? Yes.

Required? No.

You do **not** need a faceswapper to make the basic continuity workflow work.

What you need is:

- a good pair anchor
- a clean shot handoff
- a controlled sit-down transition

That can all be done without ReActor.

* * *

# The simplest working workflow for your exact film on 8 GB

Use the official **Wan 2.2 5B** path because the ComfyUI tutorial is specifically aimed at making that workable on **8 GB VRAM**. ([docs.comfy.org](https://docs.comfy.org/tutorials/video/wan/wan2_2))

Then do this:

* * *

## Phase 1 — create three stills first

Make three “anchor” images:

### Anchor A

Both characters walking side by side

### Anchor B

Both characters standing / arriving near the bench

### Anchor C

Both characters seated on the bench

Do **not** continue until those three stills look right.

That is the casting stage.

* * *

## Phase 2 — Shot 1: walk

Use **I2V** from **Anchor A**.

Keep it short.

Goal:

- small body motion
- small camera motion
- same faces
- same outfits

At the end, save the **best final frame**.

* * *

## Phase 3 — repair the bridge frame

Now inspect the last frame from Shot 1.

If the faces slipped, fix that frame **as a still**.

This is where people often think they need ReActor.

You do not.

You can repair one bridge frame:

- with a still-image workflow
- with inpainting
- or even outside ComfyUI if necessary

For your project, the last-frame repair is just a way of giving the next shot a cleaner starting point.

* * *

## Phase 4 — Shot 2: approach bench

Use the repaired last frame from Shot 1 as the input image.

Again use **I2V**.

Goal:

- slight slowing down
- slight repositioning near the bench
- no actual sit yet

At the end, save the best frame where they are standing near the bench.

* * *

## Phase 5 — Shot 3: sit

This is where I would stop using plain I2V and switch to **first/last-frame**.

Use:

- first frame = best standing-near-bench frame
- last frame = **Anchor C** (seated)

The official FLF tutorial is exactly for this kind of controlled transition. ([docs.comfy.org](https://docs.comfy.org/tutorials/video/wan/wan-flf))

That gives the model much less freedom to “recast” the pair.

* * *

## Phase 6 — Shot 4: seated

Now return to **I2V** , using the good seated frame as input.

This shot should have only gentle motion:

- breathing
- small glance
- subtle posture change

Bench shots are where you can safely create duration.

* * *

# What to do if a face drifts

Keep the fix small.

Use this rule:

## If only one key frame is bad

Repair the still, then continue.

## If a short span is bad

Use a short repair tool or short inpaint run.

## If the whole clip is bad

Regenerate that shot from the earlier trusted frame.

Do **not** try to rescue a completely broken shot by piling on more LoRAs.

* * *

# One more important quality fix

If you see weird flicker or color blinking, do not assume it is all identity drift.

There is a documented ComfyUI issue where Wan output shows **blinking/discoloration** with **VAE Decode (Tiled)**, while standard **VAE Decode** does not show the same effect. ([github.com](https://github.com/Comfy-Org/ComfyUI/issues/7116))

So for finals, if memory allows:

- use **standard VAE Decode**
- keep tiled decode only for rough previews

That can remove one whole layer of “why do my people look different?” confusion.

* * *

# My recommendation about LoRAs right now

Very simple:

## Use fewer LoRAs, not more

For your current stage, I would choose one of these:

- no LoRA
- one utility LoRA for faster previews
- one carefully tested style/identity LoRA

I would **not** stack several unknown LoRAs until your workflow is stable.

If you want faster previews, **Wan2.2-Lightning** is worth looking at because it speeds generation up a lot. That is useful for testing handoffs and anchor strategies. ([huggingface.co](https://huggingface.co/lightx2v/Wan2.2-Lightning))

But again: Lightning helps **speed** , not identity memory.

* * *

# My bottom line

You do **not** need to wait for a better GPU before making progress.

You **do** need to simplify.

The real path forward is:

1. fix the custom-node install structure
2. stop relying on Manager status messages alone
3. stop treating ReActor as mandatory
4. build the film from anchors + I2V + FLF
5. repair only the frames that matter

That is how I would make your workflow work **right now** on 8 GB.

---

<div class="post-metadata">

**Author:** ![Lance-Smith](https://avatars.discourse-cdn.com/v4/letter/l/e36b37/32.png) [@Lance-Smith](https://discuss.huggingface.co/u/Lance-Smith)\
**Post date:** [September 14, 2026, 7:52pm UTC](https://discuss.huggingface.co/t/ive-hit-a-wall-please-help/175450/3 "2026-09-14T19:52:26Z")

</div>

hmmm…

my AI imaging experice took place with a graphics card with only 8GB of VRAM.  
the important componet of an operation like this is not to throway your ideas, but to simplify your targets.  
give yourself 1 thing at a time to adjust and improve on.  
that alows you to take away winds and losses from the process. what you learn from target one can then inform target 2. and then target 3, etc.  
so as a basica stratigy. start with still images, and work on how you describe images, and then grade how well the model is following the image description. why? becase there are atleast 2 different possible failiers here:

1. bad image description (or prompt)
2. model instability.

its possible all of the things your downloaded could help EVENTUALLY. but atm you have so many working componets in your workflow that its hard to judge what is imediately usefull information, and what information cant be deciphered yet. and because of that, its almost impossible for anyone to give instructions on your direct setup on how to fix it, because knowing the componets isnt the same as understanding the exact failiers.