@movmixai

Using a fixed reference image to isolate character drift in AI videos: input management and quality assurance design

  • AI
  • generative AI
  • Python
0
Posted at 2026-11-09

When working with the same character across multiple AI video clips, looking only at differences in the output makes it impossible to determine which change to the reference image, prompt, template, or model affected the results.

In this article, we organize a method for managing inputs and tracking differences, rather than a method that completely guarantees character identity. We do not discuss performance comparisons or empirically measured success rates. The shots, filenames, and setting values used in the examples are for explanatory purposes only.

Conflict of interest disclosure: The author is a member of the Movmix team. This method does not depend on any particular product's API or internal specifications. The Python in this article is for local input management and does not connect to or send data to a video generation service.

1. Text alone does not uniquely specify a character

There is more than one character image that satisfies the description “short hair, a dark-colored shirt, and a white collar.” Even if the wording is fixed, it cannot completely define the shape of the face or the clothing.

Adding an image to the inputs provides, at minimum, an appearance that can serve as a baseline for comparison. However, a reference image is not a constraint that guarantees the correct result. When generating a profile view or back view that is not shown in the image, unseen information must be filled in. Differences between models and instructions on presentation can also affect the results.

Here, we divide the process into the following two stages.

  1. Input consistency: Whether the same image and recorded settings were used.
  2. Output conformance: Whether the generated character retains the intended appearance.

Even when the inputs are the same, the outputs are not necessarily the same. The starting point is to avoid conflating these two things.

2. Treat the reference image as an immutable input asset

Once you have chosen a baseline image, establish a workflow in which it is not silently overwritten midway through a shot.

project/ references/ character_a_v1.png prompts/ shot_01_v1.txt shot_02_v1.txt outputs/ shot_log.csv

When selecting a reference image, check the following.

  • Whether a single subject can be clearly identified
  • Whether the face, hairstyle, and costume to be maintained can be discerned
  • Whether important information has been lost due to strong shadows or occlusion
  • Whether you have permission to use the material and, where necessary, the subject's consent

If an image needs to be corrected, make it a separate file as v2. Even if you manage multiple images, do not assume that the service can accept multiple images simultaneously. Match the input range allowed on the actual screen.

Keep the hash as well as the filename

An incident in which an image with the same name was replaced cannot be detected from the filename record alone. You can obtain a SHA-256 hash using the Python standard library.

from hashlib import sha256 from pathlib import Path import json reference = Path("references/character_a_v1.png") prompt = Path("prompts/shot_01_v1.txt") record = { "shot_id": "shot_01", "reference_file": str(reference), "reference_sha256": sha256(reference.read_bytes()).hexdigest(),"prompt_file": str(prompt), "prompt_sha256": sha256(prompt.read_bytes()).hexdigest(), } print(json.dumps(record, ensure_ascii=False, indent=2))

Run this at the project root against the two files you have actually prepared. SHA-256 is used to compare the file byte sequences. The value changes even with recompression or metadata changes, and it is not an indicator for determining whether a person looks the same. Also, local records alone cannot prove what processing was performed on the service side.

3. Separate fixed conditions from shot-specific differences

If you rewrite the descriptors of appearance each time during a comparison, you will no longer be able to tell which changes were intentional.

Divide them into the following two categories.

Fixed conditions

  • Reference image file and version
  • Checklist for clothing and hairstyle to be treated as the same person
  • Selected model and template
  • Settings that can be specified in the UI, such as aspect ratio

Shot-specific changes

  • Action
  • Background
  • Framing and camera movement

For the explanatory prompt, start with a small action using a fixed camera.

The person in the reference image stands in front of a workbench and slowly turns their gaze to the right. Chest-up framing. The camera is fixed. The background is a simple indoor setting.

In the next shot, the same reference image is used, and only the motion is changed to “return the gaze to the front.” This is not a claim that it will necessarily produce a good result, but an input example in which the point of change can be tracked.

If the template contains substantial movement, we also check whether it conflicts with the “static” description in the text. If the conditions being compared cannot be aligned, that difference itself is recorded in the log.

4. Match UI feature names with the objectives to be tested

Rather than trying every product feature, select the input route that corresponds to the objective. Using names confirmed in the Movmix documentation as examples, they can be organized as follows:

Feature name Position in this task Text to Video A route that starts from text. Do not assume that a specific person’s appearance can be uniquely specified using text alone Image to Video A route for exploring video generation starting from an image Reference to Video The subject of the current investigation, using a reference image of the person Video Extend An option when considering a continuation of the same scene. It does not guarantee correction of character drift

Even when using Movmix’s Reference to Video, check the available inputs and settings on the operation screen. There are also Video to Video, AI Motion Control, and AI Recast, but adding them simultaneously while organizing the baseline conditions increases the number of changing factors.

The materials list Kling, Wan, Seedance, MiniMax, GPT Images, and Nano Banana as models. However, they are not all treated as video models for the same purpose, and the available features and models depend on the plan and the screen. This procedure does not assume an API, a fixed seed, or non-public parameters.

5. Also record the “reason for the change” in the shot log

The minimum set of columns is as follows:

shot_id,reference_file,reference_sha256,prompt_file,prompt_sha256,model,template,settings_note,output_file,decision,reason

In each field, write only the value that was actually selected or generated. There is no need to fill in items whose settings are not displayed by guessing.

If the person’s appearance drifts, limit each comparison to one major change.

  1. Keep a record of the same reference image and settings.
  2. Change only one thing, such as reducing the motion or fixing the camera.
  3. Compare the outputs using the same criteria.
  4. Record whether the output is adopted, held, or rejected, along with the reason.

A difference between a single pair of generated outputs is not enough to distinguish the effect of a change from variability during generation. If additional comparisons are added, conduct them within the available time and credit limits, and record the number of attempts. This article does not present an optimal number of attempts or success rate.

6. Divide the inspection into the beginning, during the motion, and the end

Even if the opening frames look similar, the same appearance does not necessarily persist through rotations or occlusions. Check the beginning, the middle of the motion, and the end of each clip.

Criterion What to check PersonDo the facial contours, hairstyle, and clothing structure differ significantly from the reference? Action Can the relationship between the hands and the object, the posture, and the requested action be discerned? Continuity Do the colors, positions, and orientations shift unnaturally between consecutive shots?

This is a framework for manual verification. It does not imply numerical benchmarks or identity-verification accuracy.

One point that requires particular care is when a frame from the generated video is used as the next reference image. Changes included in the previous output may carry over unchanged as the baseline. If the baseline is changed from the original image to a new one, keep them as separate versions and record the reason for the change.

Checklist

    • Finalized the reference image and avoided overwriting it
    • Recorded the image and prompt files and their hashes
    • Separated the fixed conditions from the differences between shots
    • There are no contradictions between the template and the written instructions
    • No settings that do not appear on the actual screen have been assumed
    • Compared the opening, the action sequence, and the ending with the reference image
    • Assessed the consistency of the input and the consistency of the person in the output separately
    • Recorded the reason for not using it and one item to change next

A reference image is an input asset for sharing the criteria used to make judgments, rather than a key that guarantees the same person will always be generated. By making the input traceable and fixing the perspective of inspection, this can become a production process that can at least explain “what was changed and what happened.”